Adaptive logic gradient descent control method and system for dynamic energy system parameter identification and optimization

Abstract

The application relates to a dynamic energy system parameter identification and optimization adaptive logic gradient descent control method and system. The embedded controller and the physical energy equipment unit are cooperated to collect equipment operation state data, construct a parameterized model and a target function, calculate adaptive adjustment parameters based on gradient information and normalized reference values, weight and fuse direction information of the target function and gradient direction information to generate a parameter update direction, and realize iterative optimization of a parameter vector; the system comprises a real-time measurement and sampling module, a parameterized system model unit, a target function evaluation module, a gradient calculation module, an adaptive parameter calculation module, an update direction generation module, a parameter update execution module, a convergence monitoring module and a parameter transmission and execution interface, and the convergence speed, stability and adaptability to a complex nonlinear system are improved. The application has the advantages of strong adaptability, high convergence efficiency and good real-time performance.

Description

Technical Field

[0001] This application relates to the field of power system automation and energy optimization control, and in particular to an adaptive logic gradient descent control method and system for dynamic energy system parameter identification and optimization. Background Technology

[0002] With the large-scale integration of renewable energy and the rapid development of power electronics technology, dynamic energy systems such as power electronic converters, microgrid energy management systems, and battery energy storage systems are placing higher demands on the real-time performance, stability, and adaptability of parameter identification and optimized control. In these systems, it is typically necessary to optimize and dynamically adjust system parameters online under complex operating conditions and nonlinear constraints to ensure system efficiency and safety.

[0003] In existing technologies, gradient descent, as a fundamental optimization algorithm, is widely used in power electronics control and energy management. It optimizes system performance by iteratively updating parameters along the gradient of the objective function. To improve algorithm performance, various improvement methods have been proposed, such as momentum-based gradient descent and adaptive learning rate methods (e.g., Adam, RMSProp). These methods adjust the parameter update process by introducing gradient history information or statistical features, thereby improving convergence speed and stability to some extent.

[0004] However, the aforementioned existing technologies still have the following shortcomings when applied to the real-time control of dynamic energy systems: First, existing adaptive optimization methods mostly rely on first- or second-order statistical information of the gradient to adjust the update step size element by element, lacking the ability to globally perceive the overall optimization process and making it difficult to effectively control the system according to its current convergence stage; second, existing methods usually only use gradient information to determine the update direction, without fully utilizing the global optimization information contained in the objective function value, leading to slow convergence or stagnation in flat regions or near saddle points of the objective function; third, existing methods lack a continuous and smooth adaptive switching mechanism between global search and local convergence, usually relying on fixed strategies or empirical parameters, making it difficult to balance convergence efficiency and stability; in addition, existing optimization algorithms are mostly derived from general machine learning scenarios, and their convergence analysis is mainly based on ideal mathematical models, lacking targeted engineering applicability verification and formal convergence guarantees for embedded energy systems with physical constraints and real-time requirements.

[0005] Therefore, in order to meet the application requirements of dynamic energy systems in embedded real-time control scenarios, there is an urgent need for a parameter identification and optimization method that can take into account global optimization stage perception, objective function information utilization, adaptive adjustment capability, and engineering feasibility, so as to overcome the shortcomings of existing technologies. Summary of the Invention

[0006] To improve the parameter identification and optimization of dynamic energy systems in embedded real-time control scenarios, taking into account global optimization stage perception, objective function information utilization, adaptive adjustment capability, and engineering feasibility, this application provides an adaptive logic gradient descent control method and system for parameter identification and optimization of dynamic energy systems.

[0007] The adaptive logical gradient descent control method for dynamic energy system parameter identification and optimization provided in this application adopts the following technical solution: In a first aspect, an adaptive logic gradient descent control method for dynamic energy system parameter identification and optimization is provided, executed by an embedded controller communicating with a physical energy device unit, comprising the following steps: S1: Initialize the parameter vector to be optimized, as well as the learning rate, scaling factor, sensitivity parameter, convergence threshold, and norm order; S2: Calculate the gradient information corresponding to the initial objective function based on the parameter vector, and obtain its gradient norm as a normalization reference value; S3: According to a preset sampling interval, the operating status data is collected by sensors deployed on the physical energy device unit, and the operating status data includes at least one of voltage, current, power or temperature parameters; S4: Construct an objective function based on the deviation between the running status data and the output of the parameterized model, and calculate the corresponding objective function value; S5: Calculate the gradient of the current parameter vector according to the objective function and obtain the corresponding gradient norm; S6: Calculate the adaptive adjustment parameter based on the gradient norm and the normalized reference value to characterize the degree of change in the current optimization stage; S7: Based on the adaptive adjustment parameters, the direction information corresponding to the objective function value and the gradient direction information are weighted and fused to generate the parameter update direction; S8: Iteratively update the parameter vector according to the parameter update direction and learning rate to obtain a new parameter vector; S9: Determine whether the updated gradient norm satisfies the preset convergence condition. If it does, proceed to step S10; otherwise, return to step S3 to continue iterating. S10: The optimized parameter vector is transmitted to the physical energy device unit through the hardware communication interface to realize the adjustment and control of the device's operating status.

[0008] By adopting the above technical solution, the operating status data of the physical energy equipment unit is collected in real time, and an objective function is constructed in combination with a parameterized model. Adaptive adjustment parameters are calculated using gradient information and normalized reference values, so that the parameter update process can dynamically adjust the step size and direction according to the current optimization stage. At the same time, by fusing the global error information reflected by the objective function value with the gradient direction information, the global search capability and local convergence accuracy of the algorithm under complex nonlinear conditions are effectively improved. After the convergence condition is met, the optimization results are fed back to the equipment in real time, realizing closed-loop regulation and control of the system's operating status.

[0009] Secondly, an adaptive logical gradient descent control system for dynamic energy system parameter identification and optimization includes: Physical energy equipment unit; The real-time measurement and sampling module is used to acquire the operating status data of the physical energy device unit according to a preset sampling interval. The operating status data includes voltage, current, power or temperature parameters. A parameterized system model unit is used to construct and store the parameterized mathematical model of the physical energy device unit; The objective function evaluation module is used to construct an objective function based on the deviation between the running status data and the model output results, and to calculate the corresponding error value; The gradient calculation module is used to calculate the gradient of the current model parameters according to the objective function in order to obtain the direction of parameter optimization. The adaptive parameter calculation module is used to adaptively adjust the step size parameter according to preset logical rules; An update direction generation module is used to generate parameter update directions based on the gradient information and adaptive parameters. The parameter update execution module is used to iteratively update the model parameters according to the update direction; The convergence monitoring module is used to determine whether the parameter optimization process has converged according to the preset convergence criteria and output the optimal parameters. The parameter transmission and execution interface is used to transmit optimized parameters to the physical energy equipment unit through industrial communication protocols, so as to realize the adjustment and control of the equipment's operating status.

[0010] By adopting the above technical solution, an integrated parameter identification and adaptive optimization control system for dynamic energy systems is constructed. The system acquires equipment operating status data through a real-time measurement and sampling module, which is then modeled and analyzed by a parameterized system model unit. Combined with an objective function evaluation module, the system's operating error is quantitatively characterized. Furthermore, a gradient calculation module and an adaptive parameter calculation module are used to dynamically adjust the optimization step size and direction, enabling the parameter update process to adapt to different operating conditions. An update direction generation module and a parameter update execution module work together to complete iterative parameter optimization. A convergence monitoring module performs real-time judgment on the optimization process to ensure stable algorithm convergence. Finally, the optimization results are fed back to the physical energy equipment unit through a parameter transmission and execution interface, achieving closed-loop control.

[0011] Optionally, in steps S7 to S8, a momentum update mechanism is introduced to correct the parameter update direction. Specifically, during the current iteration parameter update process, the update direction of the previous iteration is weighted and superimposed according to a preset momentum coefficient to form a new update direction. The momentum coefficient ranges from 0 to 1 and is used to improve the parameter convergence speed and reduce oscillations.

[0012] By adopting the above technical solution, a momentum update mechanism is introduced to smoothly correct the parameter update direction, enabling the iteration process to effectively utilize historical update information, reduce oscillations and accelerate convergence, thereby improving the algorithm's stability and optimization efficiency.

[0013] Optionally, in step S6, the sensitivity adjustment parameter γ in the adaptive parameter calculation module is set in the range of 5 to 100, preferably 20, to adjust the degree of influence of gradient change on step size update, thereby achieving a balance between convergence speed and stability.

[0014] By adopting the above technical solution, the range of values ​​for the sensitivity adjustment parameter γ is limited, making the adaptive step size more reasonable in response to gradient changes. This ensures convergence speed while suppressing oscillations, thereby improving algorithm stability and control accuracy.

[0015] Optionally, in step S8, a sparsity constraint mechanism is introduced by adding an L1 norm regularization term based on the parameter vector to the objective function and using a proximal operator during the parameter update process to achieve sparsity of the parameter vector, thereby reducing model complexity and improving noise resistance.

[0016] By adopting the above technical solutions, introducing L1 norm regularization and proximal operator processing, the parameter vector exhibits sparse characteristics, reducing redundant parameters, improving model simplicity and noise resistance, thereby enhancing system robustness and operating efficiency.

[0017] Optionally, based on the sparse constraints, an acceleration momentum mechanism is introduced to correct the parameter update process by constructing an acceleration sequence based on historical iteration information, thereby improving the convergence speed.

[0018] By adopting the above technical solution, an acceleration momentum mechanism is introduced to make forward corrections to parameter updates, making full use of historical iteration information, improving convergence speed and reducing the number of iterations, thereby improving optimization efficiency and real-time control performance.

[0019] Optionally, in step S5, the gradient norm is calculated using the L1 norm to enhance the algorithm's robustness to outliers; The gradient calculation module adopts a fractional gradient calculation method, which uses the Grünwald-Letnikov method or the Caputo method to perform fractional derivatives of the objective function, thereby improving the ability to model complex nonlinear systems.

[0020] By adopting the above technical solutions, the gradient norm is calculated using the L1 norm, which reduces the impact of outliers on the optimization process and improves the robustness of the algorithm. At the same time, the fractional gradient calculation method is introduced to enhance the ability to characterize the dynamic characteristics of complex nonlinear systems, thereby improving the accuracy of parameter identification and the overall optimization effect.

[0021] Optionally, the physical energy device unit is a power electronic converter, the parameter vector includes at least one or more of the following: equivalent series resistance, inductance, capacitance, or switching loss coefficient, and the objective function is constructed based on the mean square error between the actual measured value and the model predicted value of the converter output. The physical energy device unit is a microgrid energy management system. The parameter vector is used to characterize the scheduling settings of active power and reactive power. The objective function comprehensively considers power generation cost, bus voltage deviation and power allocation error.

[0022] By adopting the above technical solution, the method can perform parameter modeling and optimized control for different application scenarios such as power electronic converters and microgrid energy management systems. By constructing an objective function based on mean square error and multi-objective comprehensive index, it can achieve precise optimization of equipment parameters and power scheduling, thereby improving system operating efficiency, control accuracy and economy.

[0023] Optionally, steps S3 to S10 are executed within a single control cycle of the embedded controller, and the parameter update is ensured to be completed within a preset real-time control time window by constraining the computational complexity of the algorithm.

[0024] By adopting the above technical solution, each step is completed within a single control cycle of the embedded controller, and the complexity is calculated through a constraint algorithm to ensure that parameter updates can be completed within a preset real-time window, thereby meeting the real-time and stability requirements of the dynamic energy system and improving control response speed and engineering feasibility.

[0025] Optionally, the system may include multiple distributed local controllers, each of which corresponds to a distributed energy unit and independently executes the method of claim 1 based solely on local measurement data, without any interaction of gradient information or model parameters between the controllers.

[0026] By adopting the above technical solutions, each distributed energy unit can independently complete parameter optimization and control decisions based on local data without the need for information interaction between controllers, thereby reducing communication burden and system coupling, and improving system reliability, scalability and fault resistance.

[0027] In summary, this application includes at least one of the following beneficial technical effects: Achieving adaptive phase transition control improves convergence efficiency: By introducing an adaptive adjustment mechanism based on logic functions, the optimization process can smoothly transition between the global search stage and the local fine optimization stage without the need for preset learning rate scheduling or decay strategies, thereby significantly reducing the number of iterations and improving the convergence speed; combined with the momentum expansion mechanism, the convergence efficiency can be further improved and oscillations reduced.

[0028] By fusing objective function information to achieve global optimization awareness, the scalar information of the objective function is fused with gradient direction information to construct a two-component update direction. This enables the optimization process to not only rely on local gradients but also utilize global error information, thereby enhancing the ability to traverse platform regions and saddle points and improving the effectiveness and stability of the optimization path.

[0029] It possesses good convergence and engineering reliability: the optimization method still meets the descent requirement under non-steady-state conditions and can converge to the optimal solution or near-optimal solution under certain conditions. Compared with existing adaptive optimization methods that rely on empirical parameter tuning, it has better theoretical support and engineering reliability, and is suitable for energy control scenarios with high safety requirements.

[0030] With low computational complexity, it is suitable for embedded real-time control: It requires only a small number of scalar parameters to be calculated and stored. Compared with adaptive algorithms that require the maintenance of multidimensional statistics, it significantly reduces computational overhead and storage requirements. It can be directly deployed in embedded controllers, digital signal processors or field-programmable gate arrays to meet real-time control requirements.

[0031] It has good scalability and applicability: by introducing sparse constraints and acceleration mechanisms, it can achieve sparsification optimization of high-dimensional parameters and reduce model complexity; at the same time, combined with fractional gradient extension, it can be applied to the modeling and control of complex nonlinear and fractional dynamic systems, thereby expanding the application scope of this invention in various dynamic energy scenarios such as power electronics, energy storage systems and microgrids. Attached Figure Description

[0032] Figure 1 This is a system architecture diagram of an embodiment of this application.

[0033] Figure 2 This is a flowchart of the A-GD algorithm according to an embodiment of this application.

[0034] Figure 3 This is a flowchart of the A-MGD algorithm according to an embodiment of this application.

[0035] Figure 4 This is a block diagram of adaptive parameter calculation in an embodiment of this application.

[0036] Figure 5 This is a parameter identification diagram of a DC-DC power converter according to an embodiment of this application. Detailed Implementation

[0037] The following is in conjunction with the appendix Figure 1-5 This application will be described in further detail.

[0038] This application discloses an adaptive logic gradient descent control method for dynamic energy system parameter identification and optimization, which is executed by communication between an embedded controller and a physical energy device unit. It aims to solve the technical problems of insufficient adaptability and lack of automatic phase change control in existing gradient optimization methods for dynamic energy system parameter identification and real-time scheduling optimization. Existing methods are difficult to achieve rapid convergence, stable breakthrough of plateau period exploration, and accurate near-optimal utilization within the execution constraints of the embedded energy management controller.

[0039] Therefore, the technical solution of this application provides the following: 1. An adaptive gradient descent method (A-GD), wherein a logic control scalar calculated from the ratio of the current gradient norm to the initial gradient norm adaptively... The parameters can automatically and smoothly control the transition from global exploration to local gradient optimization without requiring any manually defined scheduling rules or threshold parameters. 2. A two-component update mechanism that integrates the objective function scalar value and gradient vector into a single update mechanism, providing global landscape perception capabilities. This is all the previous gradient-based The optimizer does not have this feature; 3. A momentum extension variant (A-MGD) that integrates a velocity accumulation mechanism within a logistic-controlled two-component framework and achieves convergence to the true extremum through formal proof; 4. A complete system architecture for real-time parameter identification and scheduling optimization of physical energy devices—including power electronic converters, battery energy storage systems, photovoltaic inverters, and microgrid energy management controllers—can be deployed on embedded DSP and FPGA hardware without additional memory or computing resources, requiring only the standard gradient descent algorithm; 5. The extended variant supports sparse parameter identification (A-GD-ISTA / FISTA) through proximal operator integration, fractional gradient calculation for fractional-order dynamic system models, and distributed multi-agent deployment for decentralized microgrid scheduling optimization.

[0040] Reference Figure 2 , Figure 3 This includes the following steps: S1: Initialize the parameter vector to be optimized, as well as the learning rate, scaling factor, sensitivity parameter, convergence threshold, and norm order; S2: Calculate the gradient information corresponding to the initial objective function based on the parameter vector, and obtain its gradient norm as a normalization reference value; S3: According to a preset sampling interval, the operating status data is collected by sensors deployed on the physical energy device unit, and the operating status data includes at least one of voltage, current, power or temperature parameters; S4: Construct an objective function based on the deviation between the running status data and the output of the parameterized model, and calculate the corresponding objective function value; S5: Calculate the gradient of the current parameter vector according to the objective function and obtain the corresponding gradient norm; S6: Calculate the adaptive adjustment parameter based on the gradient norm and the normalized reference value to characterize the degree of change in the current optimization stage; S7: Based on the adaptive adjustment parameters, the direction information corresponding to the objective function value and the gradient direction information are weighted and fused to generate the parameter update direction; S8: Iteratively update the parameter vector according to the parameter update direction and learning rate to obtain a new parameter vector; S9: Determine whether the updated gradient norm satisfies the preset convergence condition. If it does, proceed to step S10; otherwise, return to step S3 to continue iterating. S10: The optimized parameter vector is transmitted to the physical energy device unit through the hardware communication interface to realize the adjustment and control of the device's operating status.

[0041] An adaptive logic gradient descent control system for dynamic energy system parameter identification and optimization consists of ten functional components, connected as follows: Figure 1 As shown, it includes: Component 1 - Physical Energy Device Unit (PEDU) refers to physical engineering equipment that requires parameter identification or optimization, including DC-DC power converters, battery energy storage converters, photovoltaic inverters, wind turbine controllers, or microgrid energy management systems. The PEDU receives input signals and outputs measurable signals. Component 2 - Real-time Measurement and Sampling Module (RMSM): A hardware measurement subsystem containing voltage, current and power sensors, interfaced with PEDU. The RMSM samples the output signal at sampling intervals and transmits discrete-time measurement vectors to Component 4. Component 3 - Parametric System Model Unit (PSMU): The PEDU mathematical model stored in the embedded controller's memory is parameterized by the optimization variable vector. The PSMU calculates and predicts the output based on the current parameter estimates and inputs. Component 4 - Objective Function Evaluation Module (OFEM): Evaluates the scalar objective function by calculating the mean square error between the measured output and the model prediction output; function In microgrid dispatch applications, this function represents the function that includes generation. Economic dispatch cost function including cost, bus voltage deviation penalty, and power sharing error penalty; Component 5 - Gradient Calculation Module (GCM): Approximates the gradient vector using analytical differentiation or numerical finite difference, and evaluates it. -norm; Component 6 - Adaptive Parameter Calculation Module (APCM): Evaluates the adaptive parameters of logic control. Sensitivity The phase transition rate is controlled by parameters. APCM uses the initial gradient norm at the time of the first iteration as the normalization benchmark. For example... Figure 4 As shown, the output results The initial value requirements are met; Component 7 - Dual Component Update Direction Module (DCUDM): Responsible for assembling the adaptive update direction module. Among them is dimension The unit vector is a scaling factor. DCUDM guarantees strict descent conditions: in all non- At the point of equilibrium; Component 8 - Parameter Update Execution Module (PUEM): Responsible for performing parameter vector update operations. Component 9 - Convergence Monitoring and Termination Module (CMTM): Evaluates the convergence criteria in each iteration. When the criteria are met, it outputs the optimal parameters. Vector and optimal objective value; Component 10 - Parameter Assignment and Execution Interface (PDAI): Depending on the deployment platform, it uses IEC61850 Goose Message Transmission, Modbus RTU / TCP, or CAN. The bus transmits the optimized parameter vector to the PEDU actuator.

[0042] like Figure 2 As shown, the A-GD method performs the following steps sequentially in each optimization cycle: Step S1 - Initialization: Initialize the parameters using physically feasible estimates. Vector. Set the learning rate, scaling factor (nominal value: 2), and sensitivity parameter ( Nominal value: 20), Convergence tolerance ( Nominal value, norm order, and iteration counter . ; Step S2 - Initial Gradient and Norm Storage: Calculated using component 5, and the results are stored. Stored in In component 6, it serves as the normalization reference value for all subsequent iterations; Step S3 - Real-time Measurement Acquisition: Activate component 2 to acquire measurements from the PEDU sensor during the sampling interval. vector. ; Step S4 - Objective Function Evaluation: Component 4 Based on Components 3 Measurement data and model prediction Evaluate; Step S5 - Gradient Calculation: Component 5 Calculation And determine its norm. . ; Step S6 - Adaptive Parameter Calculation: Component 6 uses the stored reference norm for calculation. Calculation. Where: (Balanced exploration-development); and: (Pure gradient development). ; Step S7: Update Direction Components: Assemble Component 7 ; Step S8 - Parameter Update: Component 8 executes ; Step S9 - Convergence Check: Component 9 is evaluated; if the conditions are met... If the above steps are not taken, proceed to step S10; otherwise: Set the parameters and return to step S3; Step S10 - Output and Scheduling: Component 10, through the configured hardware... The protocol sends data to the PEDU, and the optimization cycle is complete.

[0043] A-MGD variant - Momentum extension The A-MGD method extends A-GD by replacing steps S7-S8 with a two-step velocity accumulation procedure, such as... Figure 3 As shown. The remaining steps (S1-S6, S9-S10) are exactly the same as A-GD. The initialization process also requires the introduction of a momentum coefficient. .

[0044] Step S7-M-Speed ​​Update: Step S8-M - Momentum parameter update: A-GD-ISA Variant - Sparse Optimization Extension Under the regularization condition for constraint parameter identification (applicable to fault-sparse systems) (Model and Demand Response Sparse Microgrid Scheduling), step S8 is replaced by near-end soft threshold update: in The regularization term is given, and the proximal operator is a soft threshold function: . The A-GD-FISTA variant further introduces the Nesterov momentum sequence, achieving accelerated convergence and a higher convergence rate through parameters and . for. ; Key parameter range Application Example - DC-DC Power Supply Parameter Identification like Figure 5 In the specific embodiment shown, this system is used for parameter identification of a DC-DC power converter, as component 1 (PEDU): PEDU (Component 1): DC-DC converter with input voltage Working under duty cycle , generate output voltage and output current ; RMSM (Component 2): Voltage and current sensor acquisition And sampling is performed at sampling intervals. .

[0045] PSMU (Component 3): This converter is modeled using average state-space equations, and its parameter vectors are unknown. The equivalent series resistance is The inductance is Output Capacitor is The switching loss coefficient in the on state is .

[0046] OFEM (Component 4): The objective function is the weighted mean square error between the measured output voltage and the predicted output voltage and current, which includes weighting coefficients. ; PDAI (Component 10): The identified parameter vector The data is transmitted to the converter controller via the Modbus RTU / CAN bus, which then updates the duty cycle compensation, current controller gain, and voltage loop bandwidth accordingly.

[0047] Other application implementations of the embodiments of this application This method and system can be applied without structural modifications to: 1. Microgrid Energy Dispatch Optimization: This PEDU is a microgrid energy management system composed of distributed energy resources (DER) units. The parameter vector represents the dispatch setpoints for active and reactive power. The optimized setpoints are transmitted to the DER inverter controller via the IEC61850 goose message protocol, regardless of whether it is in grid-connected or islanded operation mode. 2. Battery Parameter Identification: This PEDU is a battery energy storage system modeled using an equivalent circuit or fractional-order electrochemical impedance model. The A-GD-ISTA variant supports sparse fault identification of battery cell arrays; 3. Neural Network Training: This method, as a weight optimization algorithm, includes network weights. With bias term, It can automatically switch from extensive exploration in the early stage of training to fine-grained weight optimization in the later stage without the need to manually set the learning rate scheme.

[0048] The working principle is as follows: After the system starts, component 2 continuously collects the output data of PEDU. Component 3 maintains the parameterized mathematical model of PEDU based on the current estimate. Component 4 calculates the scalar error between the measured output and the model prediction. Component 5 calculates the gradient of the error relative to the parameter vector. Component 6 evaluates logic control parameters (The initial value is set to 0.5 to ensure the objective function) (with gradient component weight balance), as the gradient norm decreases and gradually approaches 1, the update gradually shifts to pure gradient optimization. Component 7 constructs a two-component update direction, corresponding to a one-step minimizer of the homotopy proxy. As the number of iterations increases, this direction continuously deforms from the objective function to a local quadratic approximation. Component 8 performs parameter updates, component 9 monitors convergence, and component 10 transmits the optimization parameters to the physical device upon termination.

[0049] The key technical points of this application: The core technology of this application lies in its logic-controlled dual-component adaptive gradient update mechanism. This mechanism calculates the ratio of the current gradient norm to the initial gradient norm using the logical Sigmoid function, generating a single scalar adaptive parameter. Simultaneously regulate two fundamentally different update components: the scalar value of the objective function. (Enabling globally direction-independent exploration) and gradient vectors (enabling locally precise optimization). This mechanism achieves an automatic, smooth, and irregular exploration-optimization phase transition. With strict mathematical descent guarantees, it can be applied to real-time parameter identification and optimization of physical energy devices and energy management systems. Existing gradient-based optimizers cannot achieve this combination in a single scalar control, hardware-deployable update framework.

[0050] The technical protection points of this application are as follows: The following elements constitute the core technical protection scope of this application. All protection points must be met in order to realize the key technical advantages: Protection Point 1: Logic Adaptive Parameter Formula (Component 6, Step S6): The formula for calculating the adaptive parameters is: The control signal is the ratio of the current gradient norm to the initial gradient norm—not the cumulative gradient statistic—and the output is a scalar that guarantees the initial value.

[0051] Protection point 2: Two-component update direction (component 7, step S7): The update direction is: The objective function scalar value and the gradient vector both coexist in the update direction. They interact as control weight coefficients.

[0052] Protection Point 3: Physical Energy System Parameter Update (Component 8, Step S8) This update applies to Represents energy The device can identify physical The parameter vector of the parameters and the optimization results are transmitted to the physical actuator or model register through the industrial hardware interface (component 10).

[0053] Protection point 4: Sensitivity parameter for controlling phase transition rate (component 6). Sensitivity parameter control logic of nominal value Phase transition The rate, where larger values ​​accelerate the shift towards gradient-dominated development, and This parameter is excluding the standard learning rate. The only parameter that needs to be adjusted.

[0054] Protection point 5: Momentum extension A-MGD (component 8, steps S7-M and S8-M). Accumulated speed: Among them, the same logical control mechanism originating from protection point 1 simultaneously The direction of the two components in the velocity update is regulated, while the momentum coefficient ensures velocity stability.

[0055] Alternative technical solutions to the technical solution of this application The following alternative technical solutions achieve the same technical effect by adjusting the specific implementation plan while realizing the technical goal of logistics-driven adaptive gradient optimization for dynamic energy system parameter identification and scheduling optimization.

[0056] Alternative 1 - ℓ1-norm gradient norm variant: In another implementation, the gradient norm ratio in the adaptive parameter formula adopts the -norm ( ) rather than -norm ( Perform the calculation: This variant effectively resists the influence of abnormal gradient components in the measurement of noisy energy systems, while eliminating the step of calculating the square root of the norm, thereby reducing the computational complexity of low-precision embedded controllers, and retaining the logic phase transition mechanism.

[0057] Alternative Solution 2 - Update block diagonals by coordinates: In another implementation of high-dimensional parameter identification (e.g., high-order grid impedance model parameter identification or large-scale distributed energy dispatching problems), the adaptive update algorithm updates the coordinate block randomly selected in each iteration. This design reduces the computational complexity of each iteration to a minimum while preserving the logical phase transition mechanism and formal convergence properties, thus supporting implementation on resource-constrained embedded hardware. It can efficiently identify high-order dynamic models.

[0058] Alternative Option 3 – Random Variance Reduction Extension (A-SVRE-GD): In alternative approaches to system identification for large datasets (e.g., extracting battery aging parameters from massive historical operational data), the gradient calculation in step S5 is replaced by variance-reduced stochastic gradient estimation using SVRE techniques. This method reduces the computational complexity of each iteration to... It employs a randomized step size level while retaining the logistic regression adaptive mechanism and its phase transition characteristics, thereby enabling scalable deployment in data-intensive energy system identification processes.

[0059] Alternative Solution 4 – Fractional Gradient Extension: In an alternative implementation suitable for identifying fractional-order dynamic system models (including fractional-order battery state-of-charge models and fractional impedance spectrum models), the standard gradient in step S5 is replaced with a fractional-order gradient, which is approximated using the Grunwald-Letnikov or Caputo fractional derivative. Accordingly, the adaptive parameters in component 6 are obtained using the fractional gradient norm ratio. Calculations were performed to fully preserve the fractional-order optimized landscape. Logical phase transition mechanism. .

[0060] Alternative Solution 5 – Multi-Agent Distributed Expansion: In another implementation of distributed microgrid dispatch optimization, when multiple distributed energy devices (DERs) are involved, each DER controller runs a local adaptive gradient descent. (A-GD) instance, which has a local objective function and local gradient. Each controller communicates via... Independently compute local gradients Norm determination Adaptive parameters enable a fully decentralized optimization process without the need for gradient information exchange between devices. By adding a consensus penalty term to the local objective function, coordination constraints between DER units can be forced. This design makes distributed energy management scalable while avoiding the additional overhead of continuous inter-device communication.

[0061] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. An adaptive logic gradient descent control method for parameter identification and optimization of a dynamic energy system, characterized in that, The process, executed by the embedded controller communicating with the physical energy device unit, includes the following steps: S1: Initialize the parameter vector to be optimized, as well as the learning rate, scaling factor, sensitivity parameter, convergence threshold, and norm order; S2: Calculate the gradient information corresponding to the initial objective function based on the parameter vector, and obtain its gradient norm as a normalization reference value; S3: According to a preset sampling interval, the operating status data is collected by sensors deployed on the physical energy device unit, and the operating status data includes at least one of voltage, current, power or temperature parameters; S4: Construct an objective function based on the deviation between the running status data and the output of the parameterized model, and calculate the corresponding objective function value; S5: Calculate the gradient of the current parameter vector according to the objective function and obtain the corresponding gradient norm; S6: Calculate the adaptive adjustment parameter based on the gradient norm and the normalized reference value to characterize the degree of change in the current optimization stage; S7: Based on the adaptive adjustment parameters, the direction information corresponding to the objective function value and the gradient direction information are weighted and fused to generate the parameter update direction; S8: Iteratively update the parameter vector according to the parameter update direction and learning rate to obtain a new parameter vector; S9: Determine whether the updated gradient norm satisfies the preset convergence condition. If it does, proceed to step S10; otherwise, return to step S3 to continue iterating. S10: The optimized parameter vector is transmitted to the physical energy device unit through the hardware communication interface to realize the adjustment and control of the device's operating status.

2. An adaptive logical gradient descent control system for dynamic energy system parameter identification and optimization, characterized in that, include: Physical energy equipment unit; The real-time measurement and sampling module is used to acquire the operating status data of the physical energy device unit according to a preset sampling interval. The operating status data includes voltage, current, power or temperature parameters. A parameterized system model unit is used to construct and store the parameterized mathematical model of the physical energy device unit; The objective function evaluation module is used to construct an objective function based on the deviation between the running status data and the model output results, and to calculate the corresponding error value; The gradient calculation module is used to calculate the gradient of the current model parameters according to the objective function in order to obtain the direction of parameter optimization. The adaptive parameter calculation module is used to adaptively adjust the step size parameter according to preset logical rules; An update direction generation module is used to generate parameter update directions based on the gradient information and adaptive parameters. The parameter update execution module is used to iteratively update the model parameters according to the update direction; The convergence monitoring module is used to determine whether the parameter optimization process has converged according to the preset convergence criteria and output the optimal parameters. The parameter transmission and execution interface is used to transmit optimized parameters to the physical energy equipment unit through industrial communication protocols, so as to realize the adjustment and control of the equipment's operating status.

3. The method according to claim 1, characterized in that: In steps S7 to S8, a momentum update mechanism is introduced to correct the parameter update direction. Specifically, during the current iteration parameter update process, the update direction of the previous iteration is weighted and superimposed according to a preset momentum coefficient to form a new update direction. The momentum coefficient ranges from 0 to 1 and is used to improve the parameter convergence speed and reduce oscillations.

4. The method according to claim 1 or 3, characterized in that: In step S6, the sensitivity adjustment parameter γ in the adaptive parameter calculation module is set in the range of 5 to 100, preferably 20, to adjust the degree of influence of gradient change on step size update, thereby achieving a balance between convergence speed and stability.

5. The method according to claim 1, characterized in that: In step S8, a sparsity constraint mechanism is introduced. By adding an L1 norm regularization term based on the parameter vector to the objective function and using a proximal operator during the parameter update process, the parameter vector is made sparse, thereby reducing model complexity and improving noise resistance.

6. The method according to claim 5, characterized in that: Based on the sparse constraints, an acceleration momentum mechanism is introduced to improve the convergence speed by constructing an acceleration sequence based on historical iteration information to correct the parameter update process.

7. The method according to claim 1, characterized in that: In step S5, the gradient norm is calculated using the L1 norm to enhance the algorithm's robustness to outliers; The gradient calculation module adopts a fractional gradient calculation method, which uses the Grünwald-Letnikov method or the Caputo method to perform fractional derivatives of the objective function, thereby improving the ability to model complex nonlinear systems.

8. The method according to claim 1, characterized in that: The physical energy device unit is a power electronic converter, and the parameter vector includes at least one or more of the following: equivalent series resistance, inductance, capacitance, or switching loss coefficient. The objective function is constructed based on the mean square error between the actual measured value and the model prediction value of the converter output. The physical energy device unit is a microgrid energy management system. The parameter vector is used to characterize the scheduling settings of active power and reactive power. The objective function comprehensively considers power generation cost, bus voltage deviation and power allocation error.

9. The method according to claim 1 or 3, characterized in that: Steps S3 to S10 are executed within a single control cycle of the embedded controller, and the computational complexity of the algorithm is constrained to ensure that the parameter update is completed within the preset real-time control time window.

10. The system according to claim 2, characterized in that: It includes multiple distributed local controllers, each of which corresponds to a distributed energy unit and independently executes the method of claim 1 based solely on local measurement data, and the controllers do not interact with each other on gradient information or model parameters.

Images