A micro-grid energy storage optimal scheduling method based on predictive control

Abstract

The application provides a kind of micro-grid energy storage optimization scheduling method based on predictive control, belongs to micro-grid energy storage technical field, and the application establishes state space mathematical model containing dynamic characteristics by collecting energy storage equipment operating parameters, constructs double-layer game optimization framework to coordinate multiple physical constraints of energy storage equipment, establishes rolling time domain predictive control strategy to process dynamic response demand of energy storage equipment, establishes adaptive power regulation model based on graph convolution network to realize accurate prediction and dynamic adjustment of energy storage power, establishes energy storage equipment constraint processing mechanism to process multiple physical constraint conditions using Lagrange multiplier method, establishes prediction error compensation algorithm to monitor energy storage equipment operating state in real time and dynamically correct control instruction, executes energy storage scheduling instruction and carries out closed-loop feedback control, and solves the technical problem that energy storage equipment is difficult to realize dynamic response under multiple physical constraint conditions.

Description

Technical Field

[0001] This invention belongs to the field of microgrid energy storage technology, and more specifically, relates to a microgrid energy storage optimization scheduling method based on predictive control. Background Technology

[0002] Microgrid energy storage optimization scheduling technology is a key technology for achieving stable operation of energy storage devices in complex power grid environments. Traditional energy storage scheduling methods mainly use static mathematical models to describe the charging and discharging characteristics of energy storage devices. By establishing the state-of-charge equation and power balance constraints of the energy storage devices, traditional optimization algorithms are used to solve the charging and discharging power allocation of the energy storage devices. This approach is widely used in distributed energy storage systems, power grid peak shaving and frequency regulation, and renewable energy grid integration. However, traditional energy storage scheduling methods exhibit significant technical limitations when facing dynamic changes in the physical characteristics of energy storage devices. They cannot effectively handle the complex coupling relationship between the state of charge, temperature characteristics, and power response of energy storage devices, leading to problems such as response lag, unstable power output, and system oscillations in actual operation. In other words, existing technologies face the technical challenge of achieving dynamic response of energy storage devices under multiple physical constraints. Summary of the Invention

[0003] In view of this, the present invention provides a microgrid energy storage optimization scheduling method based on predictive control, which can solve the technical problem in the prior art that energy storage devices are difficult to achieve dynamic response under multiple physical constraints.

[0004] This invention is implemented as follows: This invention provides a microgrid energy storage optimization scheduling method based on predictive control, comprising: collecting operating parameters such as the state of charge, charging and discharging power, battery temperature, and terminal voltage of energy storage devices; establishing a state-space mathematical model including the dynamic characteristics of the energy storage device's charging and discharging process, temperature change characteristics, and state of charge change characteristics; constructing a two-layer game optimization framework, establishing an upper-layer economic optimization model with the objective of minimizing the economic cost of the microgrid and a lower-layer lifetime optimization model with the objective of minimizing the lifetime loss of the energy storage device, the two models being coupled through the energy storage charging and discharging power; and establishing a rolling time-domain predictive control strategy, in each control... During the control period, load forecast data and renewable energy output forecast results are acquired to establish a future energy storage dispatch forecast model in the time domain; an adaptive power regulation model is established, using an energy storage power forecast structure based on graph convolutional networks, and a confidence evaluation function is established to adjust the charge and discharge rate parameters of energy storage devices according to the confidence value; an energy storage device constraint processing mechanism is established, using the Lagrange multiplier method to handle equality and inequality constraints; a forecast error compensation algorithm is established to calculate the deviation between the actual operating state of the microgrid and the forecast value, and dynamically adjust the energy storage dispatch strategy according to the magnitude of the deviation and the confidence value; energy storage dispatch commands are executed and closed-loop feedback control is performed.

[0005] Specifically, the steps of the state-space mathematical model involve describing the input-output relationship of the energy storage device through state equations. The input variables include charging and discharging power commands, ambient temperature, load demand, and renewable energy output, while the output variables include the state of charge and terminal voltage of the energy storage device.

[0006] Specifically, the steps of the two-layer game optimization framework are as follows: the objective function of the upper-layer economic optimization model includes a weighted sum of the operating costs of energy storage equipment, grid interaction costs, renewable energy abandonment costs, and load loss costs; the constraints include power balance constraints, energy storage capacity constraints, and grid interaction power constraints.

[0007] The objective function of the lower-level lifetime optimization model includes the product of the cycle lifetime loss term and the calendar lifetime loss term, and the constraints include charge and discharge power constraints, state of charge constraints, and temperature constraints.

[0008] Specifically, the rolling time-domain predictive control strategy involves setting a finite time-domain length and achieving dynamic scheduling of energy storage devices by solving a finite time-domain optimization problem, thereby addressing the dynamics and predictive uncertainties of the microgrid energy storage system.

[0009] The adaptive power regulation model specifically includes multi-layer graph convolutional layers, a temporal attention mechanism, and a fully connected output layer. By learning the spatial correlation and temporal dependence between various energy storage devices in the microgrid, it achieves accurate prediction of energy storage power.

[0010] Specifically, the confidence evaluation function takes as input the prediction error variance, historical prediction accuracy, model training loss, and the deviation between the current prediction value and the historical mean, and outputs a confidence value between 0 and 1.

[0011] Specifically, the charge / discharge rate parameter adjustment of the energy storage device is as follows: when the confidence value is in the range of 0 to 0.25, the charge / discharge rate parameter is adjusted to 0.2C; when the confidence value is in the range of 0.25 to 0.5, it is adjusted to 0.5C; when the confidence value is in the range of 0.5 to 0.75, it is adjusted to 0.8C; and when the confidence value is in the range of 0.75 to 1, it is adjusted to 1.0C.

[0012] Specifically, the steps of the prediction error compensation algorithm include real-time monitoring of the microgrid's operating status, including actual load demand, actual renewable energy output, and actual energy storage device operating status, correcting the charging and discharging power commands of the energy storage devices, and reducing the impact of prediction uncertainty on system operation.

[0013] Specifically, the closed-loop feedback control steps involve sending the optimized energy storage charging and discharging power command to the energy storage converter, collecting the actual operating data of the energy storage device, and feeding the actual operating data back to the rolling time-domain predictive control strategy for rolling optimization calculation at the next moment.

[0014] Specifically, the energy storage device constraint processing mechanism involves setting power constraints, capacity constraints, charge / discharge rate constraints, and temperature constraints, and transforming the constrained optimization problem into an unconstrained optimization problem by introducing Lagrange multipliers.

[0015] Specifically, the training dataset for the adaptive power regulation model is established by collecting historical microgrid operation data under different seasons and weather conditions, constructing a graph structure relationship between energy storage devices, and dividing the time-series data according to a preset time window to form a training sample set and a validation sample set.

[0016] Specifically, the adaptive power adjustment model training employs a mean squared error loss function and an adaptive learning rate optimization algorithm, prevents overfitting through an early stopping mechanism, and uses a validation sample set to evaluate model performance and adjust hyperparameters.

[0017] The confidence level value affects the energy storage dispatch frequency, power regulation range, reserve capacity configuration and operational safety margin of the microgrid. When the confidence level is low, the system will increase the reserve capacity configuration of the energy storage equipment and reduce the energy storage dispatch frequency. When the confidence level is high, the system will reduce the reserve capacity configuration and increase the energy storage dispatch frequency.

[0018] The energy storage device's operating parameter data collection includes state of charge, charge / discharge power, battery temperature, and terminal voltage. The state of charge and terminal voltage are derived from the battery management system monitoring data of the energy storage device.

[0019] In the aforementioned two-layer game optimization framework, the operating cost of energy storage devices comes from the charging and discharging power and electricity price data of energy storage devices, the grid interaction cost comes from the power exchange and electricity price data between the microgrid and the main grid, and the renewable energy abandonment cost comes from the difference between the actual output and the predicted output of renewable energy.

[0020] Furthermore, the method also includes the step of selecting different inference parameter adjustment strategies for the adaptive power adjustment model based on the power backflow risk index. Specifically, when the power backflow risk index is in the range of (0, 0.3], an aggressive inference parameter adjustment strategy is adopted to increase the temporal attention weight coefficient in the graph convolutional network to a high-sensitivity mode to enhance the system's rapid response to power changes. When the power backflow risk index is in the range of (0.3, 0.6], a balanced inference parameter adjustment strategy is adopted to set the temporal attention weight coefficient to a standard mode to maintain a balance between prediction accuracy and response speed. When the power backflow risk index is in the range of (0.6, 1], a conservative inference parameter adjustment strategy is adopted to reduce the temporal attention weight coefficient to a low-sensitivity mode to improve the stability and security of the system operation.

[0021] This invention addresses the technical challenge of achieving dynamic response of energy storage devices under multiple physical constraints by establishing a state-space mathematical model incorporating the dynamic characteristics of energy storage equipment, constructing a rolling time-domain predictive control strategy, and establishing an adaptive power regulation model and prediction error compensation algorithm based on graph convolutional networks. The state-space mathematical model accurately reflects the physical behavior of energy storage devices by describing their dynamic characteristics during charging and discharging, temperature changes, and state of charge changes, overcoming the limitation of traditional static models in describing dynamic characteristics. The rolling time-domain predictive control strategy achieves dynamic scheduling of energy storage devices by solving a finite-time domain optimization problem in each control cycle, effectively handling the dynamic and time-varying characteristics of the system. The adaptive power regulation model learns the spatial correlation and temporal dependency between energy storage devices through graph convolutional networks and dynamically adjusts the charging and discharging rate parameters of the energy storage devices using a confidence evaluation function, significantly improving the energy storage devices' response to system disturbances and control accuracy. The prediction error compensation algorithm dynamically corrects the charging and discharging power commands of the energy storage devices by monitoring the deviation between the operating state and the predicted values ​​in real time, ensuring accurate dynamic response of the energy storage devices under multiple physical constraints. In summary, the present invention solves the technical problem mentioned in the background art that energy storage devices are difficult to achieve dynamic response under multiple physical constraints. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention.

[0023] Figure 2 This is a schematic diagram of the network structure of the energy storage power prediction model involved in the present invention.

[0024] Figure 3 This is a graph showing the change in the state of charge of the energy storage device in the embodiment.

[0025] Figure 4 This is a power response characteristic curve of the energy storage device in the embodiment. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0027] like Figure 1 The diagram shown is a flowchart of a microgrid energy storage optimization scheduling method based on predictive control provided by the present invention. This method includes the following steps:

[0028] S01. Collect operating parameters such as the state of charge, charging and discharging power, battery temperature and terminal voltage of the energy storage device, and establish a state-space mathematical model that includes the dynamic characteristics of the charging and discharging process, temperature change characteristics and state of charge change characteristics of the energy storage device. Describe the input-output relationship of the energy storage device through state equations. Input variables include charging and discharging power commands, ambient temperature, load demand and renewable energy output. Output variables include the state of charge and terminal voltage of the energy storage device.

[0029] S02. Construct a two-layer game optimization framework. Establish an upper-layer economic optimization model with the goal of minimizing the economic cost of microgrids. The objective function includes the weighted sum of the operating cost of energy storage equipment, grid interaction cost, renewable energy abandonment cost, and load loss cost. The constraints include power balance constraints, energy storage capacity constraints, and grid interaction power constraints. Establish a lower-layer lifetime optimization model with the goal of minimizing the lifetime loss of energy storage equipment. The objective function includes the product of the cycle lifetime loss term and the calendar lifetime loss term. The constraints include charge and discharge power constraints, state of charge constraints, and temperature constraints. The two models are coupled through the energy storage charge and discharge power.

[0030] S03. Establish a rolling time-domain predictive control strategy, acquire load forecast data and renewable energy output forecast results in each control cycle, set a finite time domain length, establish an energy storage scheduling prediction model in the future time domain, realize the dynamic scheduling of energy storage devices by solving the finite time domain optimization problem, and handle the dynamics of the system and the uncertainty of prediction.

[0031] S04. Establish an adaptive power regulation model, adopting an energy storage power prediction structure based on graph convolutional networks, including multi-layer graph convolutional layers, a temporal attention mechanism, and a fully connected output layer. By learning the spatial correlation and temporal dependence between various energy storage devices in the microgrid, accurate prediction of energy storage power is achieved. Establish a confidence evaluation function, with inputs including prediction error variance, historical prediction accuracy, model training loss, and the deviation between the current prediction value and the historical mean. The output is a confidence value between 0 and 1. The charging and discharging rate parameters of the energy storage devices are adjusted according to the confidence value to achieve dynamic regulation of energy storage power output.

[0032] S05. Establish a constraint handling mechanism for energy storage equipment, set power constraints to limit the maximum charging and discharging power of energy storage equipment, set capacity constraints to limit the state of charge range of energy storage equipment, set charge and discharge rate constraints to limit the charging and discharging rate of energy storage equipment, set temperature constraints to limit the operating temperature range of energy storage equipment, and use the Lagrange multiplier method to handle equality constraints and inequality constraints. By introducing the Lagrange multiplier, the constrained optimization problem is transformed into an unconstrained optimization problem.

[0033] S06. Establish a prediction error compensation algorithm to monitor the microgrid's operating status in real time, including actual load demand, actual renewable energy output, and actual energy storage device operating status. Calculate the deviation between the actual operating status of the microgrid and the predicted value. Dynamically adjust the energy storage scheduling strategy based on the magnitude of the deviation and the confidence level, correct the charging and discharging power commands of the energy storage devices, and reduce the impact of prediction uncertainty on system operation.

[0034] S07. Execute the energy storage scheduling command and perform closed-loop feedback control. Send the optimized energy storage charging and discharging power command to the energy storage converter and control the energy storage device to perform charging and discharging operations according to the command. At the same time, collect the actual operating data of the energy storage device, including actual charging and discharging power, actual state of charge, actual temperature and actual terminal voltage. Feed the actual operating data back to the rolling time domain predictive control strategy for rolling optimization calculation at the next moment.

[0035] The state-space mathematical model describes the dynamic behavior of the energy storage device. Inputs include charge / discharge power commands, ambient temperature, load demand, and renewable energy output. Outputs are the state of charge (SBC) and terminal voltage of the energy storage device. The SBC originates from battery management system monitoring data and is used for capacity constraint processing in step S05. The charge / discharge power commands originate from energy storage scheduling commands in step S07 and are used for the coupled calculation of the two-layer game optimization framework in step S02. The ambient temperature originates from the microgrid environmental monitoring system and is used as input to the adaptive power regulation model in step S04. The load demand originates from the load forecasting system and is used for the rolling time-domain predictive control strategy in step S03. The renewable energy output originates from the renewable energy forecasting system and is used for the rolling time-domain predictive control strategy in step S03. The terminal voltage originates from battery management system monitoring data and is used for voltage constraint processing in step S05.

[0036] The objective function of the upper-level economic optimization model in the two-layer game optimization framework is an economic cost function, which includes a weighted sum of energy storage device operating costs, grid interaction costs, renewable energy curtailment costs, and load shortage costs. Constraints include power balance constraints, energy storage capacity constraints, and grid interaction power constraints. The energy storage device operating cost is derived from the charging and discharging power and electricity price data of the energy storage device, and is used for the economic optimization of the upper-level economic optimization model. The grid interaction cost is derived from the power exchange and electricity price data between the microgrid and the main grid, and is used for the economic optimization of the upper-level economic optimization model. The renewable energy curtailment cost is derived from the difference between the actual and predicted output of renewable energy, and is used for the economic optimization of the upper-level economic optimization model. The load shortage cost is derived from the difference between load demand and actual power supply capacity, and is used for the economic optimization of the upper-level economic optimization model. The objective function of the lower-level lifetime optimization model is a lifetime loss function, which includes the product of cyclic lifetime loss terms and calendar lifetime loss terms. Constraints include charging and discharging power constraints, state of charge constraints, and temperature constraints. The cyclic lifetime loss term is derived from the number and depth of charge and discharge cycles of the energy storage device, and is used for lifetime optimization of the lower-level lifetime optimization model. The calendar lifetime loss term is derived from the operating time and temperature conditions of the energy storage device and is used for lifetime optimization in the lower-level lifetime optimization model.

[0037] The rolling time-domain predictive control strategy is used to handle the dynamics and uncertainties of the system. By solving a finite-time-domain optimization problem within each control cycle, dynamic scheduling of energy storage devices is achieved. The control cycle is derived from the time setting of the microgrid dispatch system and is used for closed-loop feedback control in step S07. The finite-time domain is derived from the prediction range setting of the energy storage dispatch prediction model and is used for establishing the energy storage dispatch prediction model in step S03.

[0038] The adaptive power regulation model employs a graph convolutional network-based energy storage power prediction structure, comprising multiple graph convolutional layers, a temporal attention mechanism, and a fully connected output layer. The model learns the spatial correlations and temporal dependencies between energy storage devices in the microgrid to achieve accurate prediction and regulation of energy storage power. The multiple graph convolutional layers, derived from the structural design of graph convolutional networks, are used to learn the spatial correlations between energy storage devices. The temporal attention mechanism, derived from the design of attention mechanisms, is used to learn the temporal dependencies of energy storage devices. The fully connected output layer, derived from the structural design of neural networks, is used to output the energy storage power prediction results.

[0039] The steps for establishing the training dataset for the adaptive power regulation model include collecting historical microgrid operation data under different seasonal and weather conditions, including energy storage device charging and discharging power, state of charge, load demand, renewable energy output, and environmental parameters. The data undergoes preprocessing and standardization. A graph structure relationship is constructed between energy storage devices. The time-series data is segmented according to a preset time window to form training and validation sample sets. The historical microgrid operation data originates from data records of the microgrid monitoring system and is used for training the adaptive power regulation model. The graph structure relationship is derived from the physical connections and electrical relationships of energy storage devices within the microgrid and is used for constructing the graph convolutional network. The time window is determined by the time scale setting of predictive control and is used for segmenting and processing the time-series data.

[0040] The training steps of the adaptive power adjustment model include initializing model parameters and hyperparameter settings, using a mean squared error loss function and an adaptive learning rate optimization algorithm for model training, preventing overfitting through an early stopping mechanism during training, evaluating model performance and adjusting hyperparameters using a validation sample set, and saving model parameters for online prediction and scheduling after training. The model parameters originate from the initialization of neural network weights and biases, used for model learning and optimization. The hyperparameters originate from the model training configuration settings, used to control the model's learning process. The mean squared error loss function originates from the loss function design of a regression task, used to evaluate the model's prediction accuracy. The adaptive learning rate optimization algorithm originates from the design of an optimization algorithm, used for updating model parameters. The early stopping mechanism originates from the technical design to prevent overfitting, used to control the model training process. The validation sample set originates from the partitioning of the training dataset, used for evaluating model performance.

[0041] The Lagrange multiplier method is used to handle equality and inequality constraints in energy storage scheduling optimization problems. By introducing Lagrange multipliers, it transforms constrained optimization problems into unconstrained optimization problems. The equality constraints are derived from the mathematical expression of power balance constraints and are used to ensure the power balance of microgrids. The inequality constraints are derived from the physical limitations of energy storage devices and are used to ensure the safe operation of these devices. The Lagrange multipliers are derived from the mathematical theory of the Lagrange multiplier method and are used to solve constrained optimization problems.

[0042] The confidence assessment function quantifies the reliability of the prediction results. Inputs include prediction error variance, historical prediction accuracy, model training loss, and the deviation between the current predicted value and the historical mean. The output is a confidence value between 0 and 1. The prediction error variance is derived from the statistical difference between the predicted result and the actual value, used to assess the uncertainty of the prediction. The historical prediction accuracy is derived from the statistical analysis of historical prediction results, used to assess the historical performance of the model. The model training loss is derived from the loss function value during model training, used to assess the learning effect of the model. The deviation between the current predicted value and the historical mean is derived from the statistical analysis of the predicted values, used to assess the degree of anomaly in the prediction. The confidence value is used for adjusting the charge / discharge rate parameters of the energy storage device in step S04.

[0043] When the confidence level is between 0 and 0.25, a conservative weighted adjustment function is used to adjust the charge / discharge rate parameter of the energy storage device to 0.2C to improve system stability. When the confidence level is between 0.25 and 0.5, a standard weighted adjustment function is used to adjust the charge / discharge rate parameter to 0.5C to maintain system performance. When the confidence level is between 0.5 and 0.75, an aggressive weighted adjustment function is used to adjust the charge / discharge rate parameter to 0.8C to accelerate system response. When the confidence level is between 0.75 and 1, a high-efficiency weighted adjustment function is used to adjust the charge / discharge rate parameter to 1.0C to achieve rapid response. The charge / discharge rate parameter of the energy storage device is derived from the technical specifications of the energy storage device and is used to control the charge / discharge rate of the energy storage device. The conservative weighted adjustment function is derived from the design of the confidence level assessment function and is used for parameter adjustment under low confidence conditions. The standard weighted adjustment function is derived from the design of the confidence level assessment function and is used for parameter adjustment under medium to low confidence conditions. The positive weight adjustment function is derived from the design of the confidence assessment function and is used for parameter adjustment under medium-to-high confidence conditions. The efficient weight adjustment function is also derived from the design of the confidence assessment function and is used for parameter adjustment under high confidence conditions.

[0044] Different confidence levels in the prediction error compensation algorithm directly affect the microgrid's energy storage dispatch frequency, power regulation amplitude, reserve capacity configuration, and operational safety margin. The energy storage dispatch frequency is derived from the time setting of the rolling time-domain predictive control strategy and is used to control the dispatch interval of energy storage devices. The power regulation amplitude is derived from the power output range of the energy storage devices and is used to control the power variation range of the energy storage devices. The reserve capacity configuration is derived from the capacity planning of the energy storage devices and is used to ensure the system's power supply reliability. The operational safety margin is derived from the system's safety requirements and is used to ensure the safe operation of the system. When the confidence level is low, the system increases the reserve capacity configuration of the energy storage devices and reduces the energy storage dispatch frequency to ensure power supply reliability; when the confidence level is high, the system reduces the reserve capacity configuration and increases the energy storage dispatch frequency to achieve optimal economic efficiency.

[0045] The state-space mathematical model is a mathematical expression used to describe the dynamic behavior characteristics of energy storage devices in a microgrid, including the changes in the state of charge, power response characteristics, and temperature dynamic characteristics of the energy storage devices. The two-layer game-theoretic optimization framework is a two-layer optimization structure comprising an upper-layer economic optimization model and a lower-layer lifetime optimization model, coupled through the energy storage charging and discharging power. The rolling time-domain predictive control strategy is a control method that achieves dynamic scheduling by solving a finite-time-domain optimization problem within each control cycle. The adaptive power regulation model is a power prediction and regulation model based on graph convolutional networks to learn the spatial correlation and temporal dependence of energy storage devices. The confidence evaluation function is an evaluation function used to quantify the reliability of the prediction results, outputting a confidence value between 0 and 1. The prediction error compensation algorithm is a compensation algorithm that dynamically adjusts the energy storage scheduling strategy by monitoring the deviation between the microgrid operating status and the predicted value in real time.

[0046] The specific implementation methods of the above steps are described in detail below.

[0047] The specific implementation of step S01 involves establishing a state-space mathematical model of the energy storage device. First, the state-of-charge (SOC) data of the energy storage device is acquired in real time via a data acquisition module. This data originates from the energy integral calculation in the battery management system. SOC estimation is performed using a combination of Coulomb counting and open-circuit voltage methods, with a reference threshold set between 10% and 90%. When acquiring charge / discharge power data, real-time power values ​​are obtained through a power sensor, and a filtering algorithm is used to eliminate measurement noise. The power sampling frequency is set to 1kHz. Battery temperature is acquired through a thermocouple sensor, with a temperature control range set to -10℃ to 60℃. Terminal voltage is monitored in real time through a voltage sensor, with a voltage range set to 85% to 115% of the rated voltage. Based on the acquired operating parameters, a state-space mathematical model is established. The state equations are constructed using linear time-invariant system theory. State variables include SOC and battery temperature; input variables include charge / discharge power commands and ambient temperature; and output variables include SOC and terminal voltage. The model uses a Kalman filter algorithm for state estimation, achieving an accurate description of the dynamic characteristics of the energy storage device through two steps: prediction update and measurement update.

[0048] The specific implementation of step S02 involves constructing a two-layer game optimization framework. The upper-layer economic optimization model uses a mixed-integer linear programming algorithm to solve the economic cost minimization problem. The objective function includes four components: energy storage device operating cost, grid interaction cost, renewable energy curtailment cost, and load loss cost. The energy storage device operating cost is calculated based on time-of-use pricing and charging / discharging power. The grid interaction cost is determined based on a two-way pricing mechanism. The renewable energy curtailment cost is calculated by multiplying a fixed unit price by the amount of curtailed electricity. The load loss cost is calculated by multiplying the interruption price by the amount of missing load. Power balance constraints ensure supply and demand balance, energy storage capacity constraints limit the state of charge within a safe range, and grid interaction power constraints limit power exchange with the main grid. The lower-layer lifetime optimization model uses a dynamic programming algorithm to solve the problem of minimizing energy storage device lifetime loss. The cycle lifetime loss term calculates the number and depth of charge / discharge cycles based on the rainflow counting method, and the calendar lifetime loss term calculates the impact of temperature and time on lifetime based on the Arrhenius equation. Charging / discharging power constraints limit power within the rated power range, state of charge constraints prevent overcharging and over-discharging, and temperature constraints ensure the battery operates within a reasonable temperature range. The two-layer model achieves coupling through the energy storage charging and discharging power, and uses the Nash equilibrium solution algorithm to find the game equilibrium solution.

[0049] The specific implementation of step S03 involves establishing a rolling time-domain predictive control strategy, with a control cycle set at 15 minutes and a finite time domain length set at 4 hours. At the beginning of each control cycle, load forecast data and renewable energy output forecast results are acquired. Load forecasting employs a Long Short-Term Memory (LSTM) network algorithm, based on historical load data and meteorological information, with a forecast accuracy requirement of over 85%. Renewable energy output forecasting uses numerical weather prediction combined with machine learning algorithms; wind power forecasting uses a Support Vector Machine (SVM) algorithm, and photovoltaic (PV) forecasting uses a Radial Basis Function (RBF) neural network algorithm. A future time-domain energy storage scheduling prediction model is established, using a quadratic programming algorithm to solve a finite-time-domain optimization problem. The optimization variable is the energy storage charging and discharging power for each period within the future time domain. During the rolling optimization process, only the optimization results for the current period are executed; the optimization calculation is recalculated for the next period. The predictive control strategy handles model uncertainty through a feedback correction mechanism; when the deviation between the actual operating state and the predicted value exceeds 10%, re-optimization is triggered.

[0050] The specific implementation of step S04 involves establishing an adaptive power regulation model, employing a graph convolutional network to learn the spatial correlation and temporal dependency between energy storage devices. The model contains three graph convolutional layers, each with 64 neurons, and uses a spectral domain graph convolutional algorithm to handle the spatial relationships between energy storage devices. The temporal attention mechanism uses a self-attention algorithm, calculating attention weights through three matrices: query, key-value, and value, with a temporal window length set to 24 hours. The fully connected output layer contains two hidden layers, each with 32 neurons, using a modified linear unit activation function. The confidence evaluation function is constructed based on Bayesian theory, with input parameters including prediction error variance, historical prediction accuracy, model training loss, and the deviation between the current prediction value and the historical mean. The prediction error variance is calculated using a sliding window with a window length set to 168 hours. The historical prediction accuracy is calculated using the mean absolute percentage error, with a threshold set to 15%. The model training loss uses the mean squared error loss function, with a convergence threshold set to 0.001. The deviation between the current prediction value and the historical mean is standardized using Z-scores, with an outlier threshold set to 2 standard deviations. The confidence scores were calculated by weighted summation, with weighting coefficients set to 0.3, 0.25, 0.25, and 0.2, respectively.

[0051] The specific implementation of step S05 involves establishing a constraint handling mechanism for energy storage devices. Power constraints are set based on the rated power of the energy storage device, with the upper limit for charging power set at 90% of the rated power and the upper limit for discharging power set at 95% of the rated power. Capacity constraints are set based on battery safety operation requirements, with the lower limit for state of charge set at 10% and the upper limit set at 90%. Charge / discharge rate constraints are set based on battery characteristics, with the maximum charge / discharge rate set at 1C and the normal operating rate set at 0.5C. Temperature constraints are set based on battery thermal management requirements, with the operating temperature range set at 5℃~45℃. If the temperature is too high, the charge / discharge power is reduced; if the temperature is too low, preheating is required before charging / discharging. The Lagrange multiplier method is used to handle the constraints. By introducing the Lagrange multiplier, the constrained optimization problem is transformed into an unconstrained optimization problem. Equality constraints are handled using equality Lagrange multipliers, and inequality constraints are handled using Karoš-Kuntucker conditions. The Lagrange function contains a linear combination of the original objective function and constraint terms. The necessary condition for obtaining the optimal solution is obtained by taking the partial derivative of the Lagrange function and setting it equal to zero. The interior-point method is used to solve the Lagrange optimization problem, and the iterative convergence threshold is set to 0.0001.

[0052] The specific implementation of step S06 involves establishing a prediction error compensation algorithm. A real-time monitoring module collects microgrid operating status data, with a sampling frequency set to 1Hz. Actual load demand is obtained through smart meters, and data communication utilizes low-power wide-area network (LPWAN) technology. Actual renewable energy output is obtained through a power generation equipment monitoring system, with a data accuracy requirement of over 98%. The actual operating status of energy storage equipment is obtained through an energy storage management system, including parameters such as charging / discharging power, state of charge, temperature, and voltage. Deviation calculation uses a root mean square error (RMSE) algorithm, triggering a compensation mechanism when the load prediction deviation exceeds 5%. The renewable energy output prediction deviation threshold is set at 10%, and the energy storage equipment operating status deviation threshold is set at 3%. The compensation algorithm dynamically adjusts the energy storage scheduling strategy based on the deviation magnitude and confidence level, employing a proportional-integral-derivative (PID) control algorithm for power command correction. The proportional coefficient is set to 0.8, the integral coefficient to 0.2, and the derivative coefficient to 0.1. The weight of the integral term is increased when the deviation persists, and the weight of the derivative term is increased when the deviation changes drastically. The compensation algorithm also includes an anti-integral saturation mechanism to prevent excessively large integral terms from causing system instability.

[0053] The specific implementation of step S07 involves executing energy storage scheduling commands and performing closed-loop feedback control. The optimized energy storage charging and discharging power commands are sent to the energy storage converter via a communication network. The communication protocol uses the industrial Ethernet standard, and the data transmission delay is controlled within 50 milliseconds. After receiving the power commands, the energy storage converter controls the power switching devices using pulse width modulation technology to achieve precise control of the charging and discharging power of the energy storage device. The power command execution accuracy is required to be above 95%, and the response time is required to be within 1 second. Actual operating data acquisition uses a distributed data acquisition system. The actual charging and discharging power is calculated using current and voltage sensors, with a sampling accuracy required to be within 0.5%. The actual state of charge is obtained through the battery management system, and an extended Kalman filter algorithm is used for state estimation. The actual temperature is obtained through distributed temperature sensors, which are placed at key locations on the battery module. The actual terminal voltage is obtained through a high-precision voltage measurement module, with a measurement accuracy required to be within 0.1%. The feedback control system compares the actual operating data with the set values, and automatically adjusts the control parameters when the deviation exceeds the set threshold. The feedback data is filtered and calibrated by the data preprocessing module, and then input into the rolling time-domain predictive control strategy for rolling optimization calculation at the next time step.

[0054] The adaptive power regulation model employs a deep graph convolutional network structure, comprising an input layer, multiple graph convolutional layers, a temporal attention mechanism, a fully connected layer, and an output layer. The input layer receives historical operating data from each energy storage device in the microgrid. The data dimension is calculated as the number of nodes multiplied by the time step and the feature dimension. The number of nodes corresponds to the number of energy storage devices. The time step is set to 168 hours, and the feature dimension includes eight features: charging / discharging power, state of charge, temperature, and voltage. The first graph convolutional layer uses a spectral domain graph convolution algorithm with Chebyshev multinomial approximation. It learns the spatial relationships between energy storage devices through the eigenvalue decomposition of the Laplacian matrix. The kernel size is set to 3, and the number of output channels is set to 64. The second graph convolutional layer uses a graph attention network algorithm. It adaptively learns the importance weights between nodes through an attention mechanism. The number of attention heads is set to 8, and the output dimension of each head is set to 8. The third graph convolutional layer uses a graph sampling and aggregation algorithm. It extracts local features through neighborhood sampling and information aggregation. The number of sampling neighbors is set to 10, and the aggregation function is max pooling. The temporal attention mechanism employs a multi-head self-attention algorithm, calculating temporal dependencies through query matrices, key matrices, and value matrices. The number of attention heads is set to 12, with each head having a dimension of 64. Positional encoding uses sine and cosine coding to represent the positional relationships of temporal information. The fully connected layer contains two hidden layers: the first hidden layer contains 128 neurons, and the second hidden layer contains 64 neurons. The activation function is a gated linear unit, and regularization uses dropout with a dropout rate of 0.1. The output layer contains neurons corresponding to the power prediction results for each energy storage device, and the activation function is a linear function.

[0055] The training dataset setup comprises four steps: data collection, preprocessing, graph structure construction, and sample partitioning. The data collection phase gathers microgrid operation data for different seasons: three months each in spring, summer, autumn, and winter, totaling 12 months of continuous operation data. Data is collected under various weather conditions, including typical weather types such as sunny, cloudy, overcast, rainy, and snowy days, ensuring data diversity and representativeness. The collected data includes multiple dimensions such as energy storage device charging and discharging power, state of charge, battery temperature, terminal voltage, load demand, renewable energy output, ambient temperature, humidity, and wind speed. The data preprocessing phase begins with data cleaning, removing outliers and missing values. Outlier detection employs statistical methods, setting a threshold of 3 times the standard deviation. Missing value handling uses interpolation: linear interpolation for short-term missing values ​​and seasonal decomposition interpolation for long-term missing values. Data standardization uses Z-score standardization to standardize each dimension of data to a distribution with a mean of 0 and a standard deviation of 1. In the graph structure construction phase, a graph structure is built based on the physical connection relationships of energy storage devices in the microgrid. An adjacency matrix is ​​used to represent the connections between nodes, and connection weights are calculated based on electrical distance and power transmission capacity. Electrical distance is calculated using an impedance matrix, and power transmission capacity is calculated based on line capacity and transmission loss. In the sample partitioning phase, the time-series data is divided using a sliding window, with a time window length of 168 hours and a sliding step size of 1 hour. The dataset is divided into training, validation, and test sets in an 8:1:1 ratio to ensure the continuity of the time-series data. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning and model selection, and the test set is used for final performance evaluation.

[0056] Furthermore, it includes an empirical equation for power backflow risk to quantify the system stability risk level when a microgrid transmits power to the main grid. Inputs include the state-of-charge (SOC) deviation of energy storage devices, renewable energy output volatility, load demand forecasting error variance, and grid connection point voltage stability margin. The output is a power backflow risk index value between 0 and 1. The SOC deviation of energy storage devices is derived from the difference between the actual SOC of the energy storage devices and the target SOC, used to assess the adequacy of the energy storage buffer capacity. The renewable energy output volatility is derived from the statistical analysis of the variation range between actual and predicted renewable energy output, used to assess the uncertainty of renewable energy output. The load demand forecasting error variance is derived from the statistical analysis of the difference between the load forecast results and actual demand, used to assess the impact of uncertainty on the load side. The grid connection point voltage stability margin is derived from voltage monitoring data at the connection point between the microgrid and the main grid, used to assess the stability conditions of grid connection. When the power backflow risk index is in the range of (0, 0.3), an aggressive inference parameter adjustment strategy is adopted to increase the temporal attention weight coefficient in the adaptive power regulation model to a high-sensitivity mode to enhance the system's rapid response to power changes. When the power backflow risk index is in the range of (0.3, 0.6), a balanced inference parameter adjustment strategy is adopted to set the temporal attention weight coefficient to a standard mode to maintain a balance between prediction accuracy and response speed. When the power backflow risk index is in the range of (0.6, 1), a conservative inference parameter adjustment strategy is adopted to reduce the temporal attention weight coefficient to a low-sensitivity mode to improve the stability and security of system operation.

[0057] Specifically, the aggressive inference parameter adjustment strategy is used to optimize and adjust the neural network parameters when the power backflow risk index is in the low-risk range of (0, 0.3). It enhances the system's rapid response to power changes by increasing the temporal attention weight coefficient in the adaptive power adjustment model to 1.2 times the standard value. At the same time, it increases the charge / discharge rate parameter of the energy storage device from the standard 0.5C to 0.8C to achieve rapid power adjustment. It reduces the control cycle interval of the rolling time-domain predictive control strategy from 15 minutes to 10 minutes to increase the scheduling frequency. It also reduces the reserve capacity configuration of the energy storage device from 20% to 10% to improve economic efficiency.

[0058] The balanced inference parameter adjustment strategy is used to standardize the neural network parameters when the power backflow risk index is in the medium-risk range of (0.3, 0.6). It maintains the temporal attention weight coefficient in the adaptive power adjustment model at 1.0 times the standard value to maintain a balance between prediction accuracy and response speed, keeps the charge and discharge rate parameters of the energy storage device at the standard 0.5C level to ensure stable system operation, maintains the control cycle interval of the rolling time-domain predictive control strategy at the standard setting of 15 minutes to balance scheduling efficiency and computational load, and keeps the standby capacity configuration of the energy storage device at the standard level of 15% to take into account both economy and reliability.

[0059] The conservative inference parameter adjustment strategy is used to safely adjust the neural network parameters when the power backflow risk index is in the high-risk range of (0.6,1). It reduces the system's sensitivity to noise and disturbances by lowering the timing attention weight coefficient in the adaptive power regulation model to 0.8 times the standard value, reducing the energy storage device's charge / discharge rate parameter from the standard 0.5C to 0.2C to improve operational stability and safety margin, extending the control cycle interval of the rolling time-domain predictive control strategy from 15 minutes to 20 minutes to reduce the impact of frequent scheduling on the system, and increasing the energy storage device's reserve capacity configuration from 15% to 25% to ensure power supply reliability.

[0060] These three regulation strategies form a closed-loop feedback mechanism with the input variables in the state-space mathematical model, namely the charging and discharging power command, the output value of the confidence evaluation function, and the deviation correction magnitude of the prediction error compensation algorithm, thereby achieving the organic unity and coordinated optimization of neural network inference parameters and microgrid physical operation parameters.

[0061] By dynamically assessing the risk of power backflow, adaptive adjustment of neural network inference parameters is achieved, solving the stability control challenge in the bidirectional power flow process of microgrids. Its mechanism involves automatically reducing the sensitivity of the neural network to avoid system oscillations caused by over-response when the system faces high power backflow risk, while increasing the network's response sensitivity in low-risk environments to fully leverage the economic advantages of energy storage dispatch. This risk index-based adaptive parameter adjustment strategy effectively integrates the dual objectives of system safety and economy, achieving stable and efficient operation of the microgrid energy storage system under complex operating conditions through intelligent dynamic optimization of inference parameters.

[0062] It should be noted that the key technical ideas of this invention mainly include a two-layer game optimization framework, adaptive power adjustment based on graph convolutional networks, a rolling temporal prediction control strategy, and a confidence evaluation mechanism.

[0063] The two-level game theory optimization framework achieves efficient solutions to multi-objective optimization problems by separating economic optimization and lifetime optimization into two independent optimization levels. Traditional methods typically treat economic efficiency and lifetime as different weighted terms of a single objective function, making it difficult to balance the conflicting relationship between the two objectives. This invention uses game theory to establish upper and lower level models, with the upper level focusing on minimizing economic costs and the lower level focusing on minimizing lifetime loss. A Nash equilibrium algorithm is used to find a compromise solution acceptable to both. This hierarchical optimization method can better handle the trade-offs between objectives, avoids the problem of difficult-to-determine weight coefficients in traditional weighted summation methods, and improves the rationality and practicality of the optimization results.

[0064] The adaptive power regulation model based on graph convolutional networks can effectively learn the spatial correlations and temporal dependencies among energy storage devices in a microgrid, achieving more accurate power prediction and regulation. Traditional methods typically model each energy storage device as an independent unit, neglecting the mutual influence and synergistic effects between devices. This invention uses graph convolutional networks to model energy storage devices as a graph structure, learning the spatial relationships between devices through graph convolution operations, and capturing long-term dependencies in time series using a temporal attention mechanism. This modeling method can better reflect the complexity of the microgrid system, improve the accuracy of power prediction, and provide a more reliable foundation for the coordinated control of energy storage devices.

[0065] The rolling time-domain predictive control strategy effectively handles the dynamics and predictive uncertainties of the system through finite-time domain optimization and feedback correction mechanisms. Traditional methods typically employ fixed-time-domain optimization strategies, which struggle to adapt to dynamic changes in system state and external disturbances. This invention utilizes the concept of rolling optimization, resolving the finite-time domain optimization problem within each control cycle, executing only the optimization results for the current period, and re-optimizing based on the latest information in the next period. This strategy enables timely responses to changes in system state, reduces the cumulative effect of prediction errors, and improves the robustness and adaptability of control.

[0066] The confidence assessment mechanism quantifies the reliability of prediction results, enabling adaptive adjustment of energy storage scheduling strategies. Traditional methods typically employ fixed scheduling strategies, failing to dynamically adjust control parameters based on prediction quality. This invention establishes a confidence assessment function that comprehensively considers multiple indicators such as prediction error variance, historical prediction accuracy, model training loss, and prediction bias, outputting a confidence value between 0 and 1. The charge / discharge rate parameters of the energy storage equipment are dynamically adjusted based on the confidence level, achieving a multi-level adjustment strategy from conservative to aggressive. This mechanism can employ a conservative strategy to ensure system stability when prediction uncertainty is high, and an aggressive strategy to improve economic efficiency when prediction quality is good.

[0067] The synergistic effect of these four key technological approaches forms a complete energy storage optimization and scheduling system, which has significant technological advantages compared to traditional methods. The two-layer game optimization framework provides a scientific decision-making mechanism for the system, the graph convolutional network model provides an accurate prediction basis, the rolling time-domain control strategy provides a flexible execution mechanism, and the confidence assessment mechanism provides adaptive adjustment capabilities. These four elements work together to form an intelligent energy storage scheduling system that simultaneously guarantees economy, reliability, and adaptability, achieving a technological leap from passive response to active prediction, from fixed strategies to adaptive adjustment, and from single-objective optimization to multi-objective coordination.

[0068] It should be noted that the present invention also solves the following three technical problems.

[0069] First, there is the problem of accurately modeling the multi-physics coupling characteristics of energy storage devices. Traditional energy storage modeling methods typically model the electrical, thermal, and chemical characteristics of energy storage devices separately, neglecting the coupling relationships between multiple physics fields. This results in low model accuracy and an inability to accurately reflect the actual operating characteristics of the energy storage device. This invention establishes a state-space mathematical model that uniformly describes the dynamic characteristics of the charging and discharging process, temperature changes, and state of charge changes of the energy storage device. It accurately characterizes the input-output relationship of the energy storage device through state equations. Input variables include charging and discharging power commands, ambient temperature, load demand, and renewable energy output; output variables include the state of charge and terminal voltage of the energy storage device. This achieves accurate modeling of the multi-physics coupling characteristics of energy storage devices, providing a reliable physical basis for precise control.

[0070] Second, the problem of control strategy failure for energy storage devices under time-varying constraints. Existing energy storage control methods mainly design control strategies based on fixed constraints. When the physical constraints of the energy storage device change, the control strategy cannot adaptively adjust, leading to a decline in control performance or even failure. This invention establishes an adaptive power regulation model, adopts an energy storage power prediction structure based on graph convolutional networks, learns the spatial correlation between energy storage devices through multiple graph convolutional layers, captures the temporal dependence of energy storage devices using a temporal attention mechanism, establishes a confidence evaluation function to quantify the reliability of the prediction results, and dynamically adjusts the charge and discharge rate parameters of the energy storage device according to the confidence value. This achieves adaptive control of the energy storage device under time-varying constraints, effectively solving the failure problem of traditional control strategies when facing time-varying constraints.

[0071] Third, there is the issue of control accuracy degradation caused by the accumulation of prediction errors in energy storage systems. Traditional energy storage control methods lack effective error compensation mechanisms. When prediction errors accumulate, the actual operating state of the energy storage device deviates from the expected target, severely affecting control accuracy and system stability. This invention establishes a prediction error compensation algorithm to monitor the deviation between the actual operating state of the energy storage device and the predicted value in real time, including key parameters such as actual charging and discharging power, actual state of charge, actual temperature, and actual terminal voltage. Based on the magnitude of the deviation and the confidence level, the energy storage scheduling strategy is dynamically adjusted to correct the charging and discharging power commands of the energy storage device. The actual operating data is fed back to the rolling time-domain predictive control strategy for rolling optimization calculations at the next moment, forming a complete closed-loop error compensation mechanism. This effectively suppresses the cumulative effect of prediction errors and ensures the control accuracy and long-term stable operation of the energy storage system.

[0072] Specifically, the principle of this invention is:

[0073] This invention addresses the technical problem of achieving dynamic response in energy storage devices under multiple physical constraints. Its fundamental principle lies in establishing a multi-layered dynamic control and compensation mechanism. First, the state-space mathematical model accurately describes the input-output relationship of the energy storage device through state equations, unifying the dynamic characteristics of the charging and discharging process, temperature variation characteristics, and state of charge variation characteristics within a single mathematical framework, providing an accurate physical basis for dynamic control. Second, the rolling time-domain predictive control strategy decomposes the long-term optimization problem into multiple short-term sub-problems. Within each control cycle, the optimal control strategy is solved based on the current state and future prediction information. Only the control command for the first time step is executed, and re-optimization occurs in the next control cycle. This mechanism effectively handles the dynamic characteristics and time-varying constraints of energy storage devices. Third, the adaptive power regulation model based on graph convolutional networks constructs a graph structure relationship between energy storage devices, extracts the spatial features of these devices using graph convolutional layers, and captures their temporal dependencies using a temporal attention mechanism. This achieves accurate prediction of energy storage power. The confidence evaluation function quantifies the reliability of the prediction results based on indicators such as prediction error variance and historical prediction accuracy, dynamically adjusting the charge / discharge rate parameters of the energy storage devices to enable them to adaptively adjust power output based on prediction confidence. Fourth, the prediction error compensation algorithm compares the deviation between the actual operating state of the energy storage devices and the predicted values ​​in real time. It uses the deviation information and confidence values ​​to dynamically correct the energy storage scheduling strategy, forming a closed-loop feedback control to ensure accurate dynamic response of the energy storage devices under multiple physical constraints. This multi-level dynamic control and compensation mechanism can achieve precise dynamic response while meeting the physical constraints of the energy storage devices.

[0074] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0075] The specific implementation of step S01 is to establish a state-space mathematical model of the energy storage device, and the state equations are specifically expressed as follows:

[0076] ;

[0077] ;

[0078] In the formula, The state of charge of the energy storage device at time t, ranging from 0 to 1; This refers to charging power, measured in kW. Discharge power, in kW; For charging efficiency, the value is 0.9 to 0.95; For discharge efficiency, the value is taken as 0.9 to 0.95; Battery capacity, in kWh; Battery temperature, in °C; Power loss, unit: kW; Heat capacity, expressed in J / ℃; This refers to the ambient temperature, expressed in °C. Thermal resistance, expressed in °C / W; and This represents the system noise, ranging from -0.01 to 0.01.

[0079] in, The calculation formula is:

[0080] ;

[0081] In the formula, This is the battery current, measured in amperes (A). Internal resistance as a temperature-dependent parameter, expressed in Ω. The power consumption of auxiliary equipment is expressed in kW, with a value ranging from 0.1 to 0.5 kW.

[0082] The output equation is expressed as follows:

[0083] ;

[0084] In the formula, This refers to the output voltage, measured in volts (V). This is the open-circuit voltage, measured in volts (V). Internal resistance as a temperature-dependent parameter, expressed in Ω. For noise measurement, the range is -0.1 to 0.1V.

[0085] Among them, open circuit voltage The calculation formula is:

[0086] ;

[0087] In the formula, The battery characteristic coefficients were obtained through experimental fitting. The value ranges from 3.2 to 3.6V. The value ranges from 0.8 to 1.2V. The value ranges from -0.1 to 0.1V. The value ranges from -0.05 to 0.05V.

[0088] The specific implementation of step S02 is to construct a two-layer game optimization framework, and the objective function of the upper-layer economic optimization model is specifically expressed as follows:

[0089] ;

[0090] In the formula, Total economic cost, in yuan; and These are the charging and discharging power vectors, respectively. The number of time periods; These are weighting coefficients, with values ​​of 0.25, 0.35, 0.2, and 0.2 respectively. The cost of energy storage equipment is expressed in yuan. The cost of grid interaction is expressed in yuan. Cost of renewable energy rollover, in yuan; The cost of missing load is expressed in yuan.

[0091] The calculation formulas for each cost item are as follows:

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] In the formula, This is the electricity price, expressed in yuan / kWh. The time step is expressed in hours (h). Maintenance cost coefficient, unit is yuan / kW; Rated power, in kW; and These are the purchase price and the retail price of electricity, respectively, in yuan / kWh; and These represent the power purchased and the power sold, respectively, in kW; The unit price is the cost price, expressed in yuan / kWh; This refers to abandoned power, measured in kW. The unit price for missing cost is in yuan / kWh; This represents the missing power, measured in kW.

[0097] The objective function of the lower-level lifetime optimization model is specifically expressed as follows:

[0098] ;

[0099] In the formula, Total lifespan loss; This refers to the cycle life loss term; This is a calendar lifespan depletion item.

[0100] The cycle life loss term is specifically represented as follows:

[0101] ;

[0102] In the formula, This is the cycle life factor, with a value ranging from 0.01 to 0.05; The depth of discharge ranges from 0 to 1. The depth of influence index ranges from 1.5 to 2.0. For reference temperature, the value is 25℃; The temperature effect index ranges from 0.5 to 1.0.

[0103] The calendar lifespan depletion item is specifically represented as follows:

[0104] ;

[0105] In the formula, The activation energy is expressed in J / mol and ranges from 30,000 to 50,000 J / mol. This is the gas constant, with a value of 8.314 J / (mol·K); This is absolute temperature, measured in Kelvin (K). ; The runtime is expressed in hours (h).

[0106] The specific implementation of step S03 is to establish a rolling time-domain predictive control strategy. The predictive control optimization problem is specifically expressed as follows:

[0107] ;

[0108] In the formula, The control input vector contains the charging and discharging power commands for the energy storage device; To predict the time domain length, a value of 16 is used; To control the length of the time domain, a value of 4 is used; The predicted output at time k for time k+i includes the state of charge and the terminal voltage; The reference output is the desired state of charge and terminal voltage. The reference control input is the desired charge / discharge power. and These are the output and control weight matrices, respectively. It is a diagonal matrix, with diagonal elements ranging from 1 to 10. It is a diagonal matrix, with diagonal elements ranging from 0.1 to 1.

[0109] The state prediction equation is expressed as follows:

[0110] ;

[0111] In the formula, This is the state prediction vector, which contains the predicted values ​​of state of charge and battery temperature. Here is the state transition matrix, with dimension 1. ; To control the input matrix, the dimension is... ; Let be the perturbation input matrix, with dimension . ; This is a disturbance vector, containing disturbances related to ambient temperature and load demand; This is the state vector at the current moment, which includes the current state of charge and battery temperature; This is the control input at the current moment, i.e., the charge / discharge power command.

[0112] The specific implementation of step S04 is to establish an adaptive power adjustment model, and the confidence evaluation function is specifically expressed as follows:

[0113] ;

[0114] In the formula, This is a confidence level value, ranging from 0 to 1; These are weighting coefficients, with values ​​of 0.3, 0.25, 0.25, and 0.2 respectively. The variance of the prediction error; The maximum error variance is set to 0.1. Historical prediction accuracy; For maximum accuracy, the value is 1.0; The training loss for the model; The maximum training loss is set to 0.01. This is the current predicted value; This is the historical average. This represents the historical standard deviation.

[0115] The charge / discharge rate adjustment function is specifically expressed as follows:

[0116] ;

[0117] In the formula, This refers to the charge / discharge rate parameter.

[0118] The specific implementation method of step S05 is the same as described above, and will not be repeated in detail here.

[0119] The specific implementation of step S06 is to establish a prediction error compensation algorithm, and the compensation control algorithm is specifically expressed as follows:

[0120] ;

[0121] In the formula, This is the power compensation amount, in kW; This is a proportionality coefficient, with a value of 0.8. This is the integral coefficient, with a value of 0.2; is the differential coefficient, with a value of 0.1; The prediction error is expressed in kW.

[0122] The specific calculation of prediction error is expressed as follows:

[0123] ;

[0124] In the formula, The actual power of the i-th energy storage device is expressed in kW. Let be the predicted power of the i-th energy storage device, in kW; This represents the number of energy storage devices, ranging from 1 to 50.

[0125] The specific implementation method of step S07 is the same as described above, and will not be repeated in detail here.

[0126] It should be noted that the conservative weighting adjustment function is used for parameter adjustment under low confidence conditions, and is specifically expressed as follows:

[0127] ;

[0128] In the formula, The charge / discharge rate parameters are for conservative mode, ranging from 0.15C to 0.2C. The current confidence level assessment result ranges from 0 to 0.25. This is a conservative adjustment coefficient, with a value ranging from 0.1 to 0.3. The exponential decay coefficient ranges from 2 to 5. This is the adjustment coefficient for the quadratic term, with a value ranging from 0.05 to 0.15.

[0129] The parameter acquisition method is as follows:

[0130] The results were obtained experimentally, including step 1: conducting charge and discharge tests on the energy storage device at different confidence levels and recording the system stability index; step 2: fitting the relationship curve between confidence level and stability index using the least squares method to determine the optimal adjustment coefficient.

[0131] The data was obtained experimentally, including step 1: analyzing the system response characteristics under low confidence conditions in historical operating data; and step 2: determining the exponential decay law through statistical analysis and selecting the optimal decay coefficient.

[0132] The range is 0.05 to 0.15, which was determined through system simulation.

[0133] The standard weighted adjustment function is used for parameter adjustment in cases of low to medium confidence levels, and is specifically expressed as follows:

[0134] ;

[0135] In the formula, These are the charge / discharge rate parameters in standard mode, ranging from 0.45C to 0.55C. The current confidence level is estimated to be between 0.25 and 0.5. This is the standard adjustment coefficient, with a value ranging from 0.05 to 0.15; The frequency coefficient of the sine function ranges from 0.5 to 1.5. This is the linear adjustment coefficient, with a value ranging from 0.1 to 0.3.

[0136] The parameter acquisition method is as follows:

[0137] The results were obtained experimentally, including step 1: conducting multiple energy storage device performance tests in the low to medium confidence range and recording the charging and discharging efficiency and response time; step 2: determining the sinusoidal relationship between confidence and performance indicators through regression analysis and optimizing the adjustment coefficient.

[0138] The method used is experimental, including step 1: analyzing the periodic impact of confidence level changes on system performance; step 2: determining the main frequency component and setting the frequency coefficient through Fourier transform analysis.

[0139] The range is 0.1 to 0.3, which is determined by a multi-objective optimization algorithm.

[0140] The positive weighting adjustment function is used for parameter adjustment in medium-to-high confidence scenarios, and is specifically expressed as follows:

[0141] ;

[0142] In the formula, The charge / discharge rate parameters in active mode range from 0.75C to 0.85C. The current confidence level assessment result ranges from 0.5 to 0.75. The active adjustment coefficient is set at a value of 0.08 to 0.2. is the slope coefficient of the hyperbolic tangent function, with a value ranging from 3 to 8; This is the adjustment coefficient for the power term, ranging from 0.02 to 0.08.

[0143] The parameter acquisition method is as follows:

[0144] The method used is experimental, including step 1: conducting rapid response tests on energy storage devices in the medium-to-high confidence range to monitor power point tracking accuracy and dynamic performance; step 2: determining the nonlinear mapping relationship between confidence and response performance through neural network training, and extracting the optimal adjustment coefficient.

[0145] The method used is experimental, including step 1: studying the influence mechanism of confidence level changes on system response speed; step 2: determining the optimal slope parameter of the hyperbolic tangent function through control theory analysis.

[0146] The range is 0.02 to 0.08, and it is determined by optimization using a genetic algorithm.

[0147] The efficient weighting adjustment function is used for parameter adjustment under high confidence conditions, and is specifically expressed as follows:

[0148] ;

[0149] In the formula, The charge / discharge rate parameters in high-efficiency mode range from 0.95C to 1.05C. The current confidence level assessment result ranges from 0.75 to 1.0. For efficient adjustment, the value should be between 0.02 and 0.1. This is the scaling factor for the logarithmic function, ranging from 5 to 15; This is the correction coefficient for the quadratic term, with a value ranging from 0.01 to 0.05.

[0150] The parameter acquisition method is as follows:

[0151] The results were obtained experimentally, including step 1: conducting extreme performance tests on the energy storage device under high confidence conditions and recording the maximum power output and system stability; step 2: analyzing the logarithmic relationship between confidence and extreme performance using data mining techniques to determine the optimal adjustment coefficient.

[0152] The method used is experimental, including step 1: analyzing the logarithmic growth law of system performance within the high confidence interval; step 2: determining the optimal scaling parameter of the logarithmic function through a parameter identification algorithm.

[0153] The range is 0.01 to 0.05, determined by the particle swarm optimization algorithm.

[0154] The overall weight adjustment function can be expressed as a piecewise function:

[0155] ;

[0156] This synthesis function ensures that energy storage devices can use the most suitable charge / discharge rate parameters at different confidence levels, achieving an optimal balance between system stability and response speed.

[0157] Optionally, the mathematical expression for the empirical equation of power backflow risk is: ;

[0158] in, The power backflow risk index is a dimensionless value between 0 and 1 obtained through weighted comprehensive calculation, used to quantify the degree of power backflow risk in the current microgrid. The normalized value of the state of charge deviation of the energy storage device is obtained by calculating and normalizing the difference between the actual state of charge data collected by the battery management system of the energy storage device and the target state of charge calculated by the rolling time-domain predictive control strategy. The normalized value of the renewable energy output volatility is obtained by calculating the standard deviation and normalizing it through statistical analysis of historical output data and actual monitored output data from the renewable energy forecasting system. The normalized value of the load demand forecasting error variance is obtained through statistical analysis and normalization of the error between the forecasting data of the load forecasting system and the actual load monitoring data of the microgrid. The normalized value of the voltage stability margin at the grid connection point is obtained by calculating and normalizing the deviation between the real-time voltage monitoring data of the microgrid and the rated voltage at the connection point of the main grid. The weighting coefficient for the state of charge deviation term of energy storage equipment is determined by analyzing the historical impact of the energy storage state of charge on the risk of power backflow and the importance of the energy storage buffering capacity. The weighting coefficient for the volatility term of renewable energy output is determined by analyzing the historical impact of renewable energy output uncertainty on power backflow risk and the accuracy of volatility prediction. The weighting coefficients for the load demand forecasting error variance term are determined by analyzing the historical impact of load forecasting accuracy on the risk of power backflow and the contribution of load-side uncertainty. The weighting coefficients for the voltage stability margin term at the grid connection point are determined by analyzing the historical impact of grid connection point voltage stability on power backflow risk and the importance of grid safe operation. All weighting coefficients satisfy the following conditions: The normalization conditions.

[0159] In this embodiment, the state-space mathematical model of the energy storage device describes its dynamic characteristics through differential equations. This model considers the energy conversion efficiency during charging and discharging, the thermodynamic characteristics of battery temperature changes, and the stochastic influence of system noise. Compared to traditional static models, this dynamic model can more accurately predict the operating state of the energy storage device under different operating conditions, providing a more reliable basis for optimized control and significantly improving the accuracy and stability of energy storage scheduling.

[0160] The objective function of the two-layer game-theoretic optimization framework comprehensively considers two key indicators—economic cost and lifetime loss—through a weighted summation method. The upper-layer model focuses on economic optimization, while the lower-layer model focuses on lifetime optimization. By employing game theory, it achieves an efficient solution to the multi-objective optimization problem. This framework avoids the difficulty in determining the weight coefficients in traditional single-objective optimization methods, and can extend the lifespan of energy storage equipment while ensuring economic benefits, thus achieving a balance between economy and reliability.

[0161] The performance index function of the rolling time-domain predictive control strategy realizes the dynamic scheduling of energy storage devices through a quadratic optimization problem. The predictive control algorithm can re-optimize the control strategy based on the latest system state and prediction information in each control cycle. Compared with traditional fixed-strategy control, this method can better handle the dynamics and uncertainties of the system, significantly improve the robustness and adaptability of the control system, and reduce the impact of prediction errors on system performance.

[0162] The confidence assessment function quantifies the reliability of the prediction results through a combination of exponential and linear functions. This function comprehensively considers multiple influencing factors, including prediction error variance, historical prediction accuracy, model training loss, and prediction bias. Compared to traditional fixed-parameter adjustment methods, this confidence assessment mechanism can dynamically adjust control parameters based on prediction quality. It employs a conservative strategy to ensure system stability when prediction uncertainty is high, and a proactive strategy to improve economic efficiency when prediction quality is good, thus achieving adaptive optimization of the control strategy.

[0163] The prediction error compensation algorithm employs the proportional-integral-derivative (PID) control principle. By monitoring prediction errors in real time and dynamically compensating for them, it effectively reduces the impact of prediction uncertainty on system operation. Compared to traditional open-loop control methods, this compensation algorithm can promptly correct prediction errors, improving the accuracy and real-time performance of energy storage scheduling, and significantly enhancing the system's anti-interference capability and operational stability. The mathematical expression for PID control is as follows: By rapidly responding to the current error through the proportional term, eliminating the steady-state error through the integral term, and predicting the error change trend through the differential term, precise control of energy storage scheduling is achieved.

[0164] It should be noted that the variables involved in this invention are explained in detail in Table 1.

[0165] Table 1. Variable Explanation Table

[0166]

[0167] To better understand and implement this invention, Example 2 of a specific application scenario is provided below: A technical team implements a predictive control-based energy storage optimization scheduling method in a microgrid comprising wind power generation, photovoltaic power generation, and energy storage systems. The microgrid has a total capacity of 2MW and includes four 500kW lithium iron phosphate energy storage devices. Each energy storage device has a battery capacity of 1000kWh, a rated voltage of 800V, and an operating temperature range of -10℃ to 60℃. The microgrid is configured with an 800kW wind power generation system and a 600kW photovoltaic power generation system, connected to 1.2MW of load equipment.

[0168] The technical team first established a state-space mathematical model of the energy storage device, collecting real-time operating data including state of charge, charge / discharge power, battery temperature, and terminal voltage. The charging efficiency was set to 0.92, the discharging efficiency to 0.90, the heat capacity to 15000 J / ℃, and the thermal resistance to 0.08℃ / W. The system noise range was set to -0.005 to 0.005, and the measurement noise range was set to -0.05V to 0.05V. Battery characteristic coefficients were obtained through experimental fitting. The value is 3.4V. The value is 1.0V. The value is 0.05V. The value is -0.02V.

[0169] The technical team constructed a two-layer game-theoretic optimization framework. The weight coefficients of the upper-layer economic optimization model were set to 0.25, 0.35, 0.2, and 0.2, respectively. The electricity price in the energy storage equipment operating cost adopts a time-of-use pricing mechanism: peak-hour price is 0.8 yuan / kWh, normal-hour price is 0.6 yuan / kWh, and off-peak price is 0.3 yuan / kWh. The maintenance cost coefficient is set at 0.05 yuan / kW. The electricity purchase price in the grid interaction cost is set at 0.7 yuan / kWh, and the electricity sales price is set at 0.5 yuan / kWh. The renewable energy abandonment cost is set at 0.4 yuan / kWh, and the load loss cost is set at 2.0 yuan / kWh. In the lower-layer lifetime optimization model, the cycle lifetime coefficient is set at 0.03, the depth influence index is set at 1.8, the reference temperature is set at 25℃, and the temperature influence index is set at 0.8. The activation energy is set at 40000 J / mol, and the gas constant is taken as 8.314 J / (mol·K).

[0170] The technical team established a rolling time-domain predictive control strategy, with a control cycle set at 15 minutes, a prediction time domain length of 16 periods, and a control time domain length of 4 periods. The diagonal elements of the output weight matrix Q in the performance index function were set to 5, and the diagonal elements of the control weight matrix R were set to 0.5. The state transition matrix A was a 2×2 matrix, the control input matrix B was a 2×1 matrix, and the disturbance input matrix G was a 2×2 matrix. Load forecasting employed a Long Short-Time Memory (LSTM) network algorithm, achieving a prediction accuracy of 88%. Wind power output forecasting employed a Support Vector Machine (SVM) algorithm, achieving a prediction accuracy of 82%. Photovoltaic power output forecasting employed a Radial Basis Function (RBF) neural network algorithm, achieving a prediction accuracy of 85%.

[0171] The technical team established an adaptive power regulation model, employing a graph convolutional network to learn the spatial correlations and temporal dependencies between energy storage devices. The model contains three graph convolutional layers, each with 64 neurons, and the temporal window length for the temporal attention mechanism is set to 24 hours. The fully connected output layer contains two hidden layers, each with 32 neurons. The weight coefficients of the confidence evaluation function are set to 0.3, 0.25, 0.25, and 0.2, respectively. The maximum error variance is set to 0.08, the maximum accuracy to 1.0, and the maximum training loss to 0.008. Historical prediction accuracy is calculated using the mean absolute percentage error, and the deviation between the current prediction and the historical mean is standardized using Z-scores.

[0172] As shown in Table 2, the technical team set charge / discharge rate parameters at different confidence levels:

[0173] Table 2. Relationship between confidence level and charge / discharge rate

[0174]

[0175] The technical team developed a prediction error compensation algorithm with a proportional gain of 0.8, an integral gain of 0.2, and a derivative gain of 0.1. Prediction errors are calculated using the root mean square error, and the compensation mechanism is triggered when the load prediction deviation exceeds 5%. The threshold for renewable energy output prediction deviation is set at 10%, and the threshold for energy storage device operating status deviation is set at 3%. The compensation algorithm includes an anti-integral saturation mechanism to prevent excessively large integral terms from causing system instability.

[0176] The technical team executes energy storage dispatch commands and performs closed-loop feedback control, sending the optimized energy storage charging and discharging power commands to the energy storage converter via industrial Ethernet, with data transmission latency controlled within 45 milliseconds. The energy storage converter uses pulse width modulation technology to control power switching devices, achieving a power command execution accuracy of 97% and a response time controlled within 0.8 seconds. The sampling accuracy of the actual operation data acquisition system reaches 0.4%, and the actual state of charge is estimated using an extended Kalman filter algorithm, achieving an actual terminal voltage measurement accuracy of 0.08%.

[0177] like Figure 2 As shown, the technical team monitored the state of charge (SOC) changes of the energy storage device on a typical day. From 00:00 to 06:00, the device was charging, with the SOC gradually increasing from 30% to 85%. From 06:00 to 18:00, the device switched between charging and discharging based on load demand and renewable energy output, with the SOC fluctuating between 60% and 90%. From 18:00 to 24:00, the device was primarily discharging, with the SOC decreasing from 85% to 35%. Throughout the entire process, the SOC of the energy storage device remained within a safe range, preventing overcharging and over-discharging.

[0178] As shown in Table 3, the technical team recorded the output of renewable energy under different weather conditions:

[0179] Table 3. Statistics on Renewable Energy Output under Different Weather Conditions

[0180]

[0181] like Figure 4 As shown, the technical team analyzed the power response characteristics of the energy storage device. During load surges, the energy storage device can respond to power commands within one second, quickly providing the required power. When renewable energy output fluctuates, the energy storage device can smooth power fluctuations and maintain stable system operation. The charging and discharging power of the energy storage device can be flexibly adjusted within the range of -500kW to 500kW to meet the power balance requirements of the microgrid.

[0182] During implementation, the technical team discovered that the system's prediction error compensation algorithm effectively reduces the impact of prediction uncertainty on energy storage scheduling. When wind power output prediction deviations are large, the compensation algorithm can promptly adjust the charging and discharging power of the energy storage devices to avoid system power imbalance. When load prediction deviations exceed a threshold, the compensation algorithm can dynamically correct power commands to ensure power supply reliability.

[0183] Through analysis of 30 consecutive days of operational data, the technical team discovered that the cycle life loss of the energy storage device was effectively controlled. In traditional control methods, energy storage devices frequently switch drastically between high and low charge states, resulting in significant cycle life loss. By adopting the two-layer game optimization framework of this invention, the charging and discharging strategy of the energy storage device is more rational, avoiding unnecessary deep charging and discharging and extending the device's service life.

[0184] The technical team also discovered that the adaptive power regulation model can dynamically adjust control parameters based on prediction quality. When prediction confidence is high, the system adopts an aggressive strategy to increase the charge / discharge rate of the energy storage device, achieving better economic benefits. When prediction confidence is low, the system adopts a conservative strategy to reduce the charge / discharge rate of the energy storage device, ensuring stable system operation. This adaptive adjustment mechanism significantly improves the system's robustness and adaptability.

[0185] This invention represents a significant technological advancement over traditional energy storage scheduling methods. Traditional methods typically employ single-objective optimization, making it difficult to balance the conflicting relationship between economic efficiency and equipment lifespan. This invention utilizes a two-layer game-theoretic optimization framework, coordinating upper-layer economic optimization and lower-layer lifespan optimization to achieve effective solutions for multi-objective optimization. Traditional methods often treat energy storage devices as independent units, neglecting the mutual influence between them. This invention employs graph convolutional networks to learn the spatial correlations between energy storage devices, improving the accuracy of power prediction. Traditional methods typically use fixed control strategies, making it difficult to adapt to dynamic changes in system state. This invention employs a rolling time-domain predictive control strategy, enabling dynamic adjustment of control decisions based on the latest information, improving system adaptability. Traditional methods typically use fixed control parameters, unable to adjust based on prediction quality. This invention establishes a confidence assessment mechanism, enabling dynamic adjustment of control parameters based on prediction reliability, achieving adaptive optimization of the control strategy.

[0186] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A microgrid energy storage optimization scheduling method based on predictive control, characterized in that, include: Data on the state of charge (SBC), charging and discharging power, battery temperature, and terminal voltage of energy storage devices are collected to establish a state-space mathematical model that includes the dynamic characteristics of the charging and discharging process, temperature variation characteristics, and SBC variation characteristics of the energy storage devices. A two-layer game optimization framework is constructed, establishing an upper-layer economic optimization model aimed at minimizing the economic cost of the microgrid and a lower-layer lifetime optimization model aimed at minimizing the lifetime loss of the energy storage devices. The two models are coupled through the energy storage charging and discharging power. A rolling time-domain predictive control strategy is established, acquiring load forecast data and renewable energy output forecast results in each control cycle to establish a future time-domain energy storage scheduling prediction model. An adaptive power regulation model is established, using an energy storage power prediction structure based on a graph convolutional network, establishing a confidence evaluation function, and adjusting the energy storage device charging and discharging rate parameters according to the confidence value. An energy storage device constraint processing mechanism is established, using the Lagrange multiplier method to handle equality and inequality constraints. A prediction error compensation algorithm is established to calculate the deviation between the actual operating state of the microgrid and the predicted value, and dynamically adjust the energy storage scheduling strategy according to the magnitude of the deviation and the confidence value. Execute energy storage dispatch commands and perform closed-loop feedback control; The adjustment of the energy storage device's charge / discharge rate parameters based on confidence levels includes: when the confidence level is between 0 and 0.25, a conservative weighting adjustment function is used to adjust the energy storage device's charge / discharge rate parameters to 0.2C to improve system stability; when the confidence level is between 0.25 and 0.5, a standard weighting adjustment function is used to adjust the energy storage device's charge / discharge rate parameters to 0.5C to maintain system performance; when the confidence level is between 0.5 and 0.75, an aggressive weighting adjustment function is used to adjust the energy storage device's charge / discharge rate parameters to 0.8C to accelerate system response; and when the confidence level is between 0.75 and 1, an efficient weighting adjustment function is used to adjust the energy storage device's charge / discharge rate parameters to 1.0C to achieve rapid response. The conservative weighting adjustment function is used for parameter adjustment under low confidence conditions, and is specifically expressed as follows: ； In the formula, The charge / discharge rate parameters are for conservative mode, ranging from 0.15C to 0.2C. The current confidence level assessment result ranges from 0 to 0.

25. This is a conservative adjustment coefficient, with a value ranging from 0.1 to 0.

3. The exponential decay coefficient ranges from 2 to 5. This is the adjustment coefficient for the quadratic term, with a value ranging from 0.05 to 0.15; The standard weighted adjustment function is used for parameter adjustment in cases of low to medium confidence levels, and is specifically expressed as follows: ； In the formula, These are the charge / discharge rate parameters in standard mode, ranging from 0.45C to 0.55C. The current confidence level is estimated to be between 0.25 and 0.

5. This is the standard adjustment coefficient, with a value ranging from 0.05 to 0.15; The frequency coefficient of the sine function ranges from 0.5 to 1.

5. This is the linear adjustment coefficient, with a value ranging from 0.1 to 0.3; The positive weighting adjustment function is used for parameter adjustment in medium-to-high confidence scenarios, and is specifically expressed as follows: ； In the formula, The charge / discharge rate parameters in active mode range from 0.75C to 0.85C. The current confidence level assessment result ranges from 0.5 to 0.

75. The active adjustment coefficient is set at a value of 0.08 to 0.

2. is the slope coefficient of the hyperbolic tangent function, with a value ranging from 3 to 8; This is the adjustment coefficient for the power term, ranging from 0.02 to 0.

08.

2. The microgrid energy storage optimization scheduling method based on predictive control according to claim 1, characterized in that, The steps of the state-space mathematical model are specifically to describe the input-output relationship of the energy storage device through state equations. The input variables include charging and discharging power commands, ambient temperature, load demand and renewable energy output, and the output variables include the state of charge and terminal voltage of the energy storage device.

3. The microgrid energy storage optimization scheduling method based on predictive control according to claim 2, characterized in that, The steps of the two-layer game optimization framework are as follows: the objective function of the upper-layer economic optimization model includes a weighted sum of the operating costs of energy storage equipment, grid interaction costs, renewable energy abandonment costs, and load loss costs, and the constraints include power balance constraints, energy storage capacity constraints, and grid interaction power constraints.

4. The microgrid energy storage optimization scheduling method based on predictive control according to claim 3, characterized in that, The objective function of the lower-level lifetime optimization model includes the product of the cycle lifetime loss term and the calendar lifetime loss term, and the constraints include charge and discharge power constraints, state of charge constraints, and temperature constraints.

5. The microgrid energy storage optimization scheduling method based on predictive control according to claim 4, characterized in that, The specific steps of the rolling time-domain predictive control strategy are to set a finite time-domain length, realize the dynamic scheduling of energy storage devices by solving a finite time-domain optimization problem, and handle the dynamics and predictive uncertainties of the microgrid energy storage system.

6. The microgrid energy storage optimization scheduling method based on predictive control according to claim 5, characterized in that, The adaptive power regulation model specifically includes multi-layer graph convolutional layers, a temporal attention mechanism, and a fully connected output layer. By learning the spatial correlation and temporal dependence between various energy storage devices in the microgrid, it achieves accurate prediction of energy storage power.

7. The microgrid energy storage optimization scheduling method based on predictive control according to claim 6, characterized in that, The confidence evaluation function takes as input prediction error variance, historical prediction accuracy, model training loss, and deviation between the current prediction value and the historical mean, and outputs a confidence value between 0 and 1.

8. The microgrid energy storage optimization scheduling method based on predictive control according to claim 7, characterized in that, The closed-loop feedback control steps specifically involve sending the optimized energy storage charging and discharging power command to the energy storage converter, collecting the actual operating data of the energy storage device, and feeding the actual operating data back to the rolling time-domain predictive control strategy for rolling optimization calculation at the next moment.

9. The microgrid energy storage optimization scheduling method based on predictive control according to claim 8, characterized in that, It also includes the step of selecting different inference parameter adjustment strategies for the adaptive power adjustment model based on the power backflow risk index.

Images