Energy management optimization method and system based on LED driver

Abstract

The invention relates to the technical field of LED drivers, and discloses an energy management optimization method and system based on an LED driver, and the method comprises the steps: carrying out the parameter collection and nonlinear mapping of an input end and an output end in the LED driver, and obtaining a high-dimensional feature space of the operation state of the LED driver; inputting parameters in the high-dimensional feature space into a plurality of local parameter mapping networks to execute parameter sharing to obtain a local energy feature vector; inputting the local energy feature vector into a full-connection neural network for feature processing to obtain a PWM duty ratio adjustment amount, a switching frequency adjustment amount and a target PI controller parameter adjustment amount; performing stability analysis to obtain a stability parameter control strategy; and executing double-layer model predictive control based on a stable parameter control strategy, and outputting a real-time PWM regulation and control signal. According to the invention, the response speed and the control precision of the system are improved, and full-time-domain optimization control of energy management is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of LED drivers, and in particular to an energy management optimization method and system based on LED drivers. Background Art

[0002] Traditional LED drivers still face challenges during operation, including low energy utilization, large fluctuations in power conversion efficiency, and slow response speeds. Especially in complex operating environments, factors such as input grid fluctuations, temperature variations, and load changes can significantly impact the performance and lifespan of LED drivers. Existing control methods primarily utilize fixed-parameter PI controllers and preset PWM control strategies. These methods are unable to adaptively adjust control parameters based on system operating conditions, resulting in inefficient energy management under dynamic operating conditions.

[0003] Current energy management methods for LED drivers lack consideration for the coordinated optimization of multiple system parameters, particularly overlooking the complex nonlinear relationships between input and output parameters and internal power device parameters. Traditional control methods rely on simplified system models, making it difficult to accurately characterize the dynamic characteristics of LED drivers under various operating conditions. Control accuracy and response speed are significantly reduced, particularly under large disturbances and rapid load changes. Furthermore, existing technologies fail to fully account for the impact of grid frequency fluctuations on LED driver performance and lack forward-looking control strategies to address grid instability. Summary of the Invention

[0004] The present invention provides an energy management optimization method and system based on an LED driver, which improves the system response speed and control accuracy and realizes full-time domain optimization control of energy management.

[0005] In a first aspect, the present invention provides an energy management optimization method based on an LED driver, the energy management optimization method based on the LED driver comprising: Parameters of the input and output terminals of the LED driver are collected and nonlinearly mapped to obtain the high-dimensional feature space of the LED driver's operating status; Inputting the parameters in the high-dimensional feature space into multiple local parameter mapping networks to perform parameter sharing to obtain a local energy feature vector; Inputting the local energy feature vector into a fully connected neural network for feature processing to obtain a PWM duty cycle adjustment amount, a switching frequency adjustment amount, and a target PI controller parameter adjustment amount; Performing stability analysis on the PWM duty cycle adjustment amount, the switching frequency adjustment amount, and the target PI controller parameter adjustment amount to obtain a stable parameter control strategy; Based on the stable parameter control strategy, a two-layer model predictive control is executed to output a real-time PWM control signal.

[0006] In a second aspect, the present invention provides an energy management optimization system based on an LED driver, the energy management optimization system based on the LED driver comprising: A parameter acquisition module is used to collect parameters and perform nonlinear mapping on the input and output ends of the LED driver to obtain a high-dimensional feature space of the LED driver's operating status; A parameter sharing module, configured to input the parameters in the high-dimensional feature space into a plurality of local parameter mapping networks to perform parameter sharing and obtain a local energy feature vector; a feature processing module, configured to input the local energy feature vector into a fully connected neural network for feature processing to obtain a PWM duty cycle adjustment, a switching frequency adjustment, and a target PI controller parameter adjustment; a stability analysis module, configured to perform stability analysis on the PWM duty cycle adjustment amount, the switching frequency adjustment amount, and the target PI controller parameter adjustment amount to obtain a stable parameter control strategy; The predictive control module is used to perform double-layer model predictive control based on the stable parameter control strategy and output a real-time PWM control signal.

[0007] In the technical solution provided by the present invention, by constructing a high-dimensional feature space of the operating state of the LED driver, comprehensive perception of the input, output, and internal power device parameters is achieved, effectively capturing the complex nonlinear characteristics of the system operating state. Multiple local parameter mapping networks are used to implement a parameter sharing mechanism, which significantly improves parameter processing efficiency, reduces network redundancy, and enhances the feature extraction capability of similar physical quantities, achieving more accurate local energy characteristic characterization. Feature processing is performed through a fully connected neural network, achieving coordinated optimization of the PWM duty cycle, switching frequency, and PI controller parameters. Compared with the traditional single parameter adjustment method, the system response speed and control accuracy are significantly improved. A stability analysis mechanism is introduced, based on the stability discriminant function of the state space model, to ensure that all control parameter adjustments meet the system stability constraints, effectively avoiding system instability problems that may occur during the optimization process. A two-layer model predictive control structure is adopted to combine long-term energy efficiency optimization with short-term real-time control. It not only takes into account long-term factors such as grid frequency changes, but also can quickly respond to load changes, achieving full-time domain optimization control of energy management. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0009] Figure 1 Schematic diagram of the steps of the energy management optimization method based on the LED driver in an embodiment of the present invention; Figure 2 Schematic diagram of the structure of an energy management optimization system based on an LED driver in an embodiment of the present invention. DETAILED DESCRIPTION

[0010] Embodiments of the present invention provide an energy management optimization method and system based on an LED driver. The terms "first," "second," "third," "fourth," and so on (if any) in the description and claims of the present invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including," "having," and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units expressly listed, but may include other steps or units that are not expressly listed or that are inherent to such processes, methods, products, or apparatuses.

[0011] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 In one embodiment of the present invention, an energy management optimization method based on an LED driver includes: Step S1: performing parameter acquisition and nonlinear mapping on the input and output ends of the LED driver to obtain a high-dimensional feature space of the operating state of the LED driver; It is understandable that the execution subject of the present invention can be an energy management optimization system based on an LED driver, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0012] Specifically, multi-source parameter information is obtained from multiple key nodes of the LED driver, including the input end, output end and internal power device. At the input end, the input voltage value, input current value and driver body temperature value are collected. These parameters reflect the external power supply environment and the working status of the driver body. At the output end, the output voltage value, output current value and LED lamp group temperature value are collected to obtain the energy consumption output status and thermal characteristic changes at the load end. At the same time, the temperature status of key power devices inside the LED driver, such as power switching tubes, inductors and rectifier bridges, is collected in real time. These parameters reflect factors such as power conversion efficiency, heat generation and operating load within the system. The above three types of parameters are recorded as the input end original parameter set, the output end original parameter set and the internal power device original parameter set, respectively, and merged through a unified data processing module to form the LED driver original operating parameter set. Each parameter in the original operating parameter set is nonlinearly mapped. The nonlinear mapping method employed is based on neural networks, radial basis functions, support vector kernels, or other high-dimensional embedding algorithms. This mapping transforms the original physical parameters, which exhibit linear or weakly nonlinear relationships, into a high-dimensional feature space. This effectively represents information such as the coupling relationships between different parameters, nonlinear change trends, and sensitivities, thereby constructing discriminative static feature vectors. Simultaneously, the parameter evolution process is analyzed from a temporal perspective. Adjacent sampling points within continuous time segments are selected from the original operating parameter set, and their differential values are calculated for each parameter type. This yields a dynamic parameter differential set that describes the parameter change rate, fluctuation amplitude, and trend direction. These differential values are then processed using the same or dedicated nonlinear mapping function to obtain a set of dynamic parameter feature vectors that characterize the drive's operating dynamics over a short period of time. All parameter feature vectors are combined with the dynamic feature vectors in a one-to-one correspondence or concatenation manner to construct a multidimensional, high-dimensional operating state feature space containing both static and dynamic information.

[0013] Step S2: Input the parameters in the high-dimensional feature space into multiple local parameter mapping networks to perform parameter sharing and obtain a local energy feature vector; Specifically, based on the high-dimensional feature space, its internal parameters are divided into physical dimensions and functional relevance. All parameters in this space are then reorganized into multiple local correlation parameter sets based on their sources and relevance. These include the input voltage-current-temperature parameter set, which characterizes the power supply characteristics and temperature variation trends at the input end; the output voltage-current-LED temperature parameter set, which reflects the output stability and light source temperature rise at the load end; and the power device temperature parameter set, which primarily reflects the thermal load status and lifespan of internal components of the LED driver, such as MOS transistors, inductors, and rectifier bridges. To avoid dimensional effects between different dimensions and orders of magnitude, each parameter in each local correlation parameter set is normalized and uniformly mapped to the [0, 1] or [-1, 1] interval, ensuring that the neural network maintains consistent sensitivity to changes in each parameter. Based on the partitioned normalized parameter groups, multiple local parameter mapping networks are constructed. Each network structure consists of an input layer, at least one hidden layer, and an output layer. A feedforward neural network or residual structure network is used to extract local parameter features. To improve the model's ability to consistently represent the same physical property parameters across different networks, a parameter sharing mechanism is introduced. Weight parameters for the same physical quantity are shared between adjacent or similar local networks. Specifically, if parameters with the same physical semantics, such as "temperature" or "current," coexist in multiple local networks, the initial layer weights in their corresponding neural networks are set to the same and bounded during training to ensure that these parameters have a uniform characteristic response across different contexts. This results in a parameter-sharing local mapping network system with physical constraint consistency and local sensitivity. The normalized local parameters are input into the parameter-sharing mapping network system. A forward propagation process is used to obtain the output features of each local network, namely local feature output vectors. These vectors correspond to the deep representation information of different local parameter groups under the current operating state. These local feature output vectors are nonlinearly combined to form a local energy feature vector corresponding to each local parameter set.

[0014] Step S3: Input the local energy feature vector into a fully connected neural network for feature processing to obtain a PWM duty cycle adjustment, a switching frequency adjustment, and a target PI controller parameter adjustment; Specifically, multiple local energy feature vectors undergo dimension verification and feature concatenation to form an overall input feature vector, which is then fed into the input layer of a fully connected neural network for preliminary processing. In the input layer, the local energy feature vectors are linearly combined with the input layer weight matrix and, along with a bias term, generate an input feature representation. This representation, serving as the first-stage representation of the neural network, fuses the original physical features with local coupling information. The input feature representation is then fed layer by layer into the three hidden layers of the neural network. Each hidden layer extracts and transforms the input features through alternating linear mapping and nonlinear activation functions, resulting in a more semantically expressive deep feature map. In particular, in a multi-layered architecture, increasing the receptive field and the number of nonlinear transformations helps the model capture deeper nonlinear relationships and complex characteristics between input parameters. After obtaining the deep feature map, independent output mapping channels are designed for specific control objectives. PWM duty cycle adjustment calculations are performed based on these deep feature maps. A fully connected output layer projects the deep feature maps onto the duty cycle adjustment range, yielding the desired PWM duty cycle adjustment value for the current operating state. This value is then used to precisely adjust the energy output cycle of the LED driver. Based on the same deep feature mapping structure, another mapping branch performs switching frequency adjustment calculations, generating frequency control recommendations that match the current load changes and thermal conditions. This adjustment is directly fed back to the switch driver module, achieving dual optimization of system response time and energy consumption. To improve the closed-loop response quality of the control system, parameter adjustments for the current and voltage loop PI controllers are extracted based on the deep feature mapping. The current loop PI controller parameter adjustment calculation optimizes the current closed-loop control accuracy, outputting adjustments for the proportional and integral coefficients. The voltage loop PI controller parameter adjustment calculation focuses on optimizing output voltage stability and also outputs corresponding proportional and integral parameter corrections. These two adjustments form the core regulation basis for the controller's dual-loop structure. The current loop PI parameter adjustment and the voltage loop PI parameter adjustment are structurally combined to form the target PI controller parameter adjustment.

[0015] Step S4: performing stability analysis on the PWM duty cycle adjustment amount, the switching frequency adjustment amount, and the target PI controller parameter adjustment amount to obtain a stable parameter control strategy; Specifically, a state-space model of the LED driver is constructed based on its structure and circuit operating characteristics. This model incorporates key voltage and current signals representing the system's energy transfer process as state variables, including the voltage across the input capacitor, the current in the output inductor, and the voltage across the output capacitor. Furthermore, the PWM duty cycle adjustment, switching frequency adjustment, and target PI controller parameter adjustment in the control strategy are introduced as control variables to analyze their direct impact on the system's dynamic behavior. Based on the state-space model, a system stability discriminant function is created to measure whether the system will approach equilibrium during operation. The specific design approach is based on the concept of an energy function or Lyapunov function, ensuring that the function exhibits a downward trend during operation, indicating that the system's energy is gradually dissipating rather than continuously increasing. This stability function considers the interrelationships between state variables, including the squared terms of each variable and their cross terms, to capture the dynamic coupling between the system's internal variables. The PWM duty cycle adjustment, switching frequency adjustment, and target PI controller parameter adjustments output by the neural network are substituted into the established state model. The derivative of the stability function under control is calculated through the system evolution process to determine whether the current control parameters will lead to system stability. If the stability function fails to meet the decreasing condition within certain time periods, indicating a potential for system divergence or oscillation, the control parameters are modified. By analyzing the range of the stability function, a set of mathematical constraints is extracted. Based on this, an optimization model is established, employing a quadratic programming approach to minimize the correction amplitude while meeting all stability requirements. During this optimization process, the existing control parameters are automatically adjusted based on their deviation, approaching the set of parameters that can restore stable operation with minimal change. After the modification is complete, the control parameters that meet the stability requirements are integrated with the control adjustments output by the original neural network to form a final control strategy subject to stability constraints.

[0016] Step S5: Execute the double-layer model predictive control based on the stable parameter control strategy and output a real-time PWM control signal.

[0017] Specifically, a two-layer model predictive controller (MPC) is constructed based on the control strategy derived from the stability analysis. The top-level controller is responsible for global optimization decisions and trend prediction, while the bottom-level controller performs specific PWM signal selection and real-time response control. When constructing the top-level MPC, a comprehensive predictive model is established, encompassing the electrical characteristics of the LED driver, a dynamic prediction model for grid frequency trends, and a virtual reference model for aligning the target output. The LED driver model reflects physical processes such as load response and voltage and current variations. The grid frequency model simulates fluctuations in the external power supply environment. The virtual reference model bridges the gap between target performance indicators and actual system behavior, facilitating the formation of prediction error criteria. Based on this top-level predictive model, a top-level optimization control objective is set. This control objective serves as the basis for the overall system operation and scheduling. This objective includes deviations of output voltage and output current relative to target reference values, which measure system output quality; deviations of virtual reference model parameters, which assess whether the reference model effectively tracks the set trajectory; and power command adjustment terms, which regulate the energy input rhythm of the LED driver. The optimization control objective integrates multiple operational dimensions to form a multi-objective optimization structure that balances system stability, response speed, and energy performance. Based on the top-level optimization control objective, combined with the external grid frequency prediction and necessary damping ratio constraints, the virtual reference model control coefficients and power commands in the top-level controller are optimized to minimize various deviation functions while satisfying system stability and dynamic performance constraints. This results in a set of optimized reference model parameters and revised power control commands. The optimization results are then passed to the bottom-level model predictive controller, which performs the discretized selection of actual control commands. In the bottom-level controller, a driver instantaneous state prediction model is constructed based on these optimization results. A finite control set is defined, consisting of N preset PWM waveform combinations, each corresponding to a specific duty cycle and switching pattern strategy. These combinations represent a set of specific control actions available to the controller and are executable and hardware-implementable. For each PWM waveform control option, the driver model is used to predict the future system state within a short time window. The error between the system state and the reference value is calculated, and a deviation cost function is constructed based on this error. This method evaluates the performance of all candidate control actions, forming a set of performance evaluation values. After the evaluation is completed, the control sequence with the best performance evaluation value is selected from all candidate control actions, and the first control action is extracted from the sequence as the final control instruction to generate the PWM control signal in the current control cycle.

[0018] In this embodiment, a finite control set defined within the underlying model predictive controller is systematically modeled. This finite control set consists of a set of predefined PWM duty cycle and switching frequency combinations, each corresponding to an executable switching action sequence. These combinations are mapped into a unified switching state vector, which reflects the on / off state of the power electronic device and its corresponding PWM pattern within the current cycle. Through state mapping, the control input is directly converted into a behavioral description of the driver circuit, establishing a clear relationship between the control action and the system's dynamic evolution. A recursive calculation of the system state for T sampling cycles is performed on the switching state vector. Using the switching state vector corresponding to the current system state and a specific control option as the initial condition, a time-series iteration operation is performed according to the driver's electrical model to obtain the system state evolution process for the next T time steps under the control option, forming a system state prediction sequence. The system state prediction sequence is compared with the target trajectory given by the virtual reference model, and the error between the two at each time step is calculated to construct a state tracking error index. A control switching penalty term is added to the state tracking error index to mitigate issues such as system instability, increased hardware loss, or electromagnetic interference caused by frequent switching of PWM control actions. This penalty term is compared with the previous control cycle's options. If a switch occurs, a penalty weight is introduced, thus reflecting the control stability requirement in the objective function. The state tracking error is added to the control switch penalty term to construct a comprehensive performance index. Based on this comprehensive performance index, all possible control sequences are searched and ranked to select the optimal control candidate—the control action that performs best within the current forecast horizon. A secondary stability check is performed on the optimal control candidate, evaluating whether the corresponding system state meets the requirements for stable operation. This results in a performance evaluation value for each control option.

[0019] In an embodiment of the present invention, by constructing a high-dimensional feature space for the operating state of an LED driver, comprehensive perception of the input, output, and internal power device parameters is achieved, effectively capturing the complex nonlinear characteristics of the system's operating state. Multiple local parameter mapping networks are used to implement a parameter sharing mechanism, significantly improving parameter processing efficiency, reducing network redundancy, and enhancing the ability to extract features from similar physical quantities, enabling more accurate characterization of local energy characteristics. Feature processing through a fully connected neural network enables coordinated optimization of the PWM duty cycle, switching frequency, and PI controller parameters, significantly improving system response speed and control accuracy compared to traditional single-parameter adjustment methods. A stability analysis mechanism, based on a stability discriminant function of a state-space model, ensures that all control parameter adjustments meet system stability constraints, effectively avoiding potential system instability during the optimization process. A two-layer model predictive control structure combines long-term energy efficiency optimization with short-term real-time control. This not only accounts for long-term factors such as grid frequency variations, but also rapidly responds to load changes, achieving full-time-domain optimal control of energy management.

[0020] In a specific embodiment, the process of executing step S1 may specifically include the following steps: The input voltage value, input current value and driver temperature value of the input end of the LED driver are collected to obtain the original parameter set of the input end; The output voltage value, output current value and LED lamp group temperature value of the output end of the LED driver are collected to obtain the original parameter set of the output end; Collect the temperature values of the power switch tube, inductor and rectifier bridge inside the LED driver to obtain the original parameter set of the internal power devices; The original parameter set of the input end, the original parameter set of the output end and the original parameter set of the internal power device are combined to form the original operating parameter set of the LED driver; Perform nonlinear mapping on each parameter in the original operating parameter set to obtain a parameter feature vector, perform differential value calculation and mapping on the parameters of adjacent sampling points in the original operating parameter set to obtain a parameter dynamic feature vector; The parameter feature vector and the parameter dynamic feature vector are combined to form a high-dimensional feature space of the LED driver operation status.

[0021] Specifically, the input voltage, input current, and driver temperature values at the LED driver's input terminals are collected to reflect the driver's power supply environment, input energy status, and thermal stability. The input voltage and input current determine the driver's energy consumption baseline and exhibit significant dynamic fluctuations under different load conditions. The temperature indirectly reflects the heat dissipation accumulated by the power devices on the circuit board over long periods of operation. This information is continuously collected via sensors or built-in measurement modules to form the input-side raw parameter set. Simultaneously, key operating data is collected from the LED driver's output terminals, including output voltage, output current, and the temperature of the connected LED light cluster. The output voltage and current directly correspond to the energy supply required by the load, and their variations reflect the dynamic response efficiency of the driver's output control capabilities. The temperature of the LED cluster is a key indicator of the system's energy conversion efficiency and photoelectric conversion losses. It is related to the aging rate of the LED itself and also affects the stability of luminous flux and illuminance output. This set of parameters constitutes the output-side raw parameter set. Temperature values are collected for the power switches, inductor, and rectifier bridge within the LED driver to obtain the internal power device raw parameter set. The original input parameter sets, the original output parameter sets, and the original parameter sets of the internal power devices are merged in a time-aligned and structured manner to form the original operating parameter set of the LED driver. Nonlinear mapping is performed on each parameter in the original operating parameter set. By selecting an appropriate nonlinear function or neural network structure, the original scalar parameters are mapped into a feature vector space. This exploits the hidden nonlinear coupling relationships between the parameters and enhances the expressive power of the features through a high-dimensional representation. The mapping method employs methods such as multilayer perceptrons, radial basis function networks, or high-order polynomial transformations, with different mapping paths selected for different parameters to improve parameter discrimination and information density in the high-dimensional space. Through this mapping process, each parameter in the original operating parameter set is converted into a set of feature vectors, forming a parameter feature representation layer that reflects the internal behavioral characteristics of the driver under different operating states. The difference values of the parameters at adjacent sampling points in the original operating parameter set are calculated, and the magnitude of the change in each parameter between two consecutive time points is calculated to extract the dynamic evolution trend during system operation. These difference values effectively reflect the real-time response characteristics caused by load fluctuations, input disturbances, or control strategy adjustments. The differential data is processed through nonlinear mapping to obtain the characteristic vector of the dynamic parameters, namely the parameter dynamic characteristic vector, which reveals the driver's transient response capability and robustness under external disturbances. The parameter characteristic vector is combined with the parameter dynamic characteristic vector to form a comprehensive characteristic vector set that integrates the current state and the change trend, thus constructing a high-dimensional feature space of the LED driver's operating state.

[0022] In a specific embodiment, the process of executing step S2 may specifically include the following steps: According to the high-dimensional feature space, the input voltage-current-temperature parameter group, the output voltage-current-LED temperature parameter group, and the power device temperature parameter group are divided to obtain multiple groups of local correlation parameter sets; Normalizing the parameter values in each group of local correlation parameter sets to obtain normalized parameter values; Constructing multiple local parameter mapping networks, each of which includes an input layer, a hidden layer, and an output layer, and implementing a parameter sharing mechanism for the weight parameters of adjacent local parameter mapping networks, so that parameters related to the same physical quantity share weight values in different local parameter mapping networks, thereby obtaining a parameter-sharing local mapping network; Input the normalized parameter value into the parameter-sharing local mapping network, and obtain the local feature output vector of each local parameter mapping network through forward calculation; The local feature output vectors are nonlinearly combined to obtain the local energy feature vectors that characterize the correlation characteristics between relevant parameters.

[0023] Specifically, the high-dimensional operating state feature space is partitioned based on physical correlation. Based on the functional areas and physical effects of the parameter sources, the original high-dimensional features are divided into several localized parameter sets. The first set focuses on input parameters such as voltage, current, and driver temperature, which collectively reflect the external grid power supply stability, the driver's input response characteristics, and the thermal load background, and are categorized as the input voltage-current-temperature parameter set. The second set focuses on key data from the LED driver output side, including output voltage, output current, and LED lamp group temperature. These parameters primarily reflect the load power supply effect, electrical output stability, and the thermal variation characteristics of the lighting system, and are categorized as the output voltage-current-LED temperature parameter set. The third set focuses on the temperature state of power devices within the LED driver, such as power switches, inductors, and rectifier bridges. These devices are core components for energy conversion and signal modulation, and their temperature changes directly reflect the system's energy conversion efficiency and device reliability under high-frequency and high-voltage conditions. Therefore, they are classified as the power device temperature parameter set. The values in each parameter set are normalized. Normalization uses methods such as maximum and minimum scaling, Z-score standardization, or linear mapping based on the upper and lower limits of physical quantities to uniformly map all input parameters to the range [0, 1] or [-1, 1]. This step effectively smooths out differences in magnitude between different physical quantities, enhancing model training stability and balancing the weights of various features during learning. It also prevents a single parameter from dominating the network training process due to its excessively large absolute value, allowing features from different sources to participate in information fusion at the same scale. A corresponding local parameter mapping network is constructed for each local parameter group. Each network consists of an input layer, a hidden layer, and an output layer, and employs the basic structure of a feedforward neural network to perform nonlinear information extraction. The input layer receives the normalized local parameter vector, the hidden layer performs nonlinear feature transformation using an activation function, and the output layer further compresses or reorganizes the abstract features into an intermediate representation for use in subsequent fusion modules. These local mapping networks each independently process a parameter group. However, because some physical quantities (such as temperature) exist in multiple networks, a parameter sharing mechanism is introduced, using shared weights for input features with the same semantics across different networks. For example, if temperature appears as an input feature simultaneously at the input, output, and power device networks, the weight matrices of the corresponding input or hidden layers are bound to be identical. This design can effectively enhance the model's recognition consistency for similar physical quantities, improve training efficiency, and reduce the number of redundant parameters, making the overall model structure more compact. Multiple local mapping networks constructed under the parameter sharing mechanism constitute a modular system capable of extracting local structural features. The normalized parameters are input into the corresponding local networks, and the output vector of each network, i.e., the local feature output vector, is calculated through forward propagation.Each local output vector is a structurally stable block of information in vector form, encompassing the correlation patterns, changing trends, and underlying coupling logic of the original parameters. The local feature output vectors are nonlinearly combined to generate a local energy feature vector that reflects the global relationships and nonlinear interaction structures between the parameters. By introducing multi-layer perceptrons, attention mechanisms, or tensor cross networks, each local feature not only retains its own importance during the fusion process but also establishes a dynamic weighted relationship with other features. In this way, the contribution of different local features during fusion is adaptively adjusted based on the current system state, enhancing the final output vector's ability to express the driver's energy state.

[0024] In a specific embodiment, the process of executing step S3 may specifically include the following steps: The local energy feature vector is input into the input layer of the fully connected neural network for linear combination operation to obtain the input feature representation; The input feature representation is input into the three-layer hidden layer structure for feature extraction and transformation to obtain a deep feature map; Performing PWM duty cycle adjustment calculation based on the deep feature map to obtain the PWM duty cycle adjustment amount; Perform switching frequency adjustment calculation based on the deep feature map to obtain a switching frequency adjustment amount; Performing a current loop PI controller parameter adjustment calculation based on the deep feature map to obtain a current loop PI controller parameter adjustment amount, and performing a voltage loop PI controller parameter adjustment calculation based on the deep feature map to obtain a voltage loop PI controller parameter adjustment amount; The current loop PI controller parameter adjustment amount and the voltage loop PI controller parameter adjustment amount are combined into a target PI controller parameter adjustment amount.

[0025] Specifically, the local energy feature vector is input into the input layer of a fully connected neural network for linear combination. All input dimensions are multiplied by the input weight matrix and then added together with a bias term to generate an initial input feature representation. This input feature representation is then fed into a three-layer hidden layer structure for feature extraction and transformation. Each hidden layer consists of multiple neurons and uses an activation function (such as ReLU or Tanh) to perform nonlinear mapping to ensure the model's ability to handle complex control logic. In the first hidden layer, the input features are initially expanded into higher-dimensional intermediate feature vectors. The neurons in this layer learn the initial nonlinear interactions between local variables. In the second hidden layer, the feature representation further deepens the network's abstraction level, recognizing more complex parameter combinations, such as the coupling effect between input voltage and power device temperature or the nonlinear feedback of output current fluctuations on load response. In the third hidden layer, the network reconstructs and integrates the features extracted from the first two hidden layers, ensuring that the feature representation received by the output layer is highly separable and command-mapping compatible. PWM duty cycle adjustment calculations are performed based on this deep feature mapping. The deep feature vector is input into a linear transformation module through an independent fully connected output channel, which outputs a scalar value, representing the recommended PWM duty cycle adjustment for the current operating conditions. This adjustment directly affects the driver's output voltage amplitude and effective energy transfer time. Based on the deep feature vector, another independent output mapping channel calculates the switching frequency adjustment. Because the operating frequency of an LED driver is related to its conversion efficiency, EMI characteristics, and power loss, the prediction of its adjustment value must fully consider factors such as system temperature, input voltage fluctuations, and load dynamic response. The comprehensive features extracted by the deep network effectively identify the combined effects of these factors on frequency control and derive the most appropriate frequency correction recommendation for the current operating conditions. To optimize the closed-loop stability of the control system, the deep feature map outputs parameter adjustments for the LED driver's internal PI controller. Since the PI controller consists of two core control loops, the current loop and the voltage loop, which dynamically adjust the output current and output voltage, respectively, independent output channels are designed for each control loop during the parameter prediction stage. The PI parameter adjustment process for the current loop considers characteristics such as output current response speed, load change rate, and short-term current deviation. Its output includes dynamic correction values for the proportional and integral coefficients, which improve the controller's adaptability to rapid current changes. The PI parameter adjustment for the voltage loop focuses primarily on steady-state deviations and transient fluctuations in the output voltage. The corresponding proportional-integral adjustment values are used to improve the stability and anti-disturbance capability of voltage regulation. In the network structure, these two output channels are based on the same deep feature map, but their weight update strategies are independently optimized during training to ensure local optimality for each control objective.The current loop PI controller parameter adjustment amount and the voltage loop PI controller parameter adjustment amount are combined into a target PI controller parameter adjustment amount.

[0026] In a specific embodiment, the process of executing step S4 may specifically include the following steps: A state-space model is constructed based on the LED driver. The state variables of the state-space model include the input capacitor voltage, the output inductor current, and the output capacitor voltage. The control variables include the PWM duty cycle adjustment, the switching frequency adjustment, and the target PI controller parameter adjustment. A system stability discriminant function is created based on the state space model. The system stability discriminant function includes quadratic terms and cross terms of each state variable. Substitute the PWM duty cycle adjustment, switching frequency adjustment, and target PI controller parameter adjustment into the state space model, calculate the derivative of the system stability discriminant function along the system trajectory, and obtain the stability function; Perform range analysis on the stability function to obtain stability constraints. Based on the stability constraints, perform quadratic programming on the control parameters that do not meet the stability requirements to obtain the control parameter adjustments that meet the stability requirements. The control parameter adjustment that satisfies stability is combined with the PWM duty cycle adjustment, the switching frequency adjustment and the target PI controller parameter adjustment to obtain a stable parameter control strategy.

[0027] Specifically, a state-space model is constructed based on the LED driver. Considering the basic topology of an LED driver, which includes an input rectifier and filter circuit, a power conversion module, and an output load regulation section, the voltage across the input capacitor, the current in the output inductor, and the voltage across the output capacitor are selected as state variables in the state-space model. These three variables reflect the transient stability of the power input, the dynamic load response capability, and the control accuracy of the output voltage, respectively. Furthermore, to ensure the model's control significance, the input terms of the control variables are clearly defined, including the adjustment of the PWM duty cycle, the switching frequency, and the target PI controller parameters. These control variables influence the system dynamics at the levels of signal modulation, switching rhythm, and internal controller gain, forming the three main lines of driver operating state adjustment. Based on these variable definitions, a state-space model of the LED driver is constructed. This model uses a set of differential equations composed of state variables to describe the dynamic evolution of the system over time. The control variables are embedded in the system matrix or input terms. This allows the model to not only reflect the essence of the energy change process in the circuit but also respond to the direct impact of different control commands on the evolution of the state variables. During the mathematical modeling process, the equivalent circuits of each LED driver module are converted into state equations describing voltage and current variations. The continuous model is then differentially approximated based on the system's discrete sampling time, thereby establishing a discrete state-space expression that can be used in the digital controller's internal calculations. A system stability discriminant function is then created based on the state-space model. To ensure that this function truly reflects whether the system is converging toward a stable state, a positive-definite function structure is designed, incorporating quadratic and cross terms of the state variables. This function quantifies the system's current "energy level" and measures the coupling effects between different state variables. For example, the squared term of the input capacitor voltage represents the system's changing trend in energy storage stability at the input, while the cross term between the output inductor current and the output capacitor voltage characterizes the coupling relationship between load response and voltage regulation. By introducing these quadratic and cross terms, a stability discriminant function is constructed that is both mathematically continuous and meaningful in engineering physics. The PWM duty cycle adjustment, switching frequency adjustment, and target PI controller parameter adjustment output by the upper-layer neural network or model predictive controller are substituted into the established state-space model to simulate the system's dynamic trajectory under the current control strategy. By calculating the derivative of the stability discriminant function along this trajectory, we obtain the stability function, which describes the direction and speed of energy changes in the system, that is, whether the system tends to equilibrium or oscillates over time. If the derivative is always negative, it means that the system is in an asymptotically stable state under the current control strategy; if the derivative is zero or positive, there is a risk of stability loss or energy amplification.To clarify the feasibility boundaries of the control parameters, a range analysis of the stability function is performed. The range of the stability function's values is determined for different control parameter combinations, and a set of strict stability constraints is extracted. These constraints restrict the control parameter adjustment range. For example, certain duty cycle adjustments must be limited to a specific range to prevent divergence in the system inductor current, or certain PI controller proportional gains must be maintained within a certain range to prevent overshoot and output voltage oscillation. Stability constraints reflect the mutual constraints between control variables and the physical limitations imposed by the LED driver's electrical structure. If, during actual operation, the control parameters provided by the current neural network output or prediction model fail to meet these stability constraints (i.e., if the derivative of the stability function becomes non-negative), a control parameter correction mechanism is activated to ensure that the system remains controllable and stable under the adjusted control strategy. To minimize the correction of the original control command while satisfying the stability constraints, an optimization problem with stability constraints is formulated and solved using a quadratic programming approach. In this optimization process, the objective function is set to minimize the sum of the squares of the corrections. The constraints are the range boundaries of the stability function and the actual controller output capability. The optimization solution results in a new set of control parameter adjustments that meet the stability conditions. The control parameter adjustments obtained through quadratic programming are structurally combined with the initial output PWM duty cycle adjustments, switching frequency adjustments, and target PI controller parameter adjustments to generate a set of stable parameter control strategies.

[0028] In a specific embodiment, the process of executing step S5 may specifically include the following steps: According to the stable parameter control strategy, a prediction model of the top-level model predictive controller in the two-layer model predictive controller is constructed. The prediction model includes an LED driver model, a grid frequency change prediction model, and a virtual reference model. Setting a top-level optimization control target for the top-level model predictive controller, the top-level optimization control target including an output voltage deviation term, an output current deviation term, a virtual reference model parameter deviation term, and a power command adjustment term; According to the top-level optimization control target and the grid frequency forecast value, combined with the damping ratio constraint condition, the virtual reference model control coefficient and power command are optimized and calculated to obtain the optimized virtual reference model control coefficient and corrected power command; Establishing a bottom-level model predictive controller based on the optimized virtual reference model control coefficients and the modified power command, and defining a finite control set of N preset PWM waveform combinations of the bottom-level model predictive controller; Perform system state prediction and deviation cost calculation for each control option in the underlying model predictive controller to obtain the performance evaluation value of each control option; The first control action of the optimal control sequence is selected based on the performance evaluation value, and a real-time PWM control signal of the LED driver is generated.

[0029] Specifically, a predictive model for the top-level model predictive controller (MPC) in a two-layer MPC is constructed based on a stable parameter control strategy. Based on the actual circuit model of the LED driver, a mathematical model of the system's behavior is established to reflect the driver's dynamic response under varying control input and load conditions. This model accounts for the impact of input voltage fluctuations on the input capacitor, the mechanism by which the PWM modulation strategy affects the output inductor current, and the feedback path of load changes on output voltage stability. Through the dynamic coupling of electrical component parameters and state variables, a set of system state evolution equations is formed that supports time-series deduction. To improve the predictive model's adaptability to actual power grid environments, a grid frequency variation prediction model is introduced. This model characterizes the impact of input disturbances caused by external power supply fluctuations and uses time series modeling to predict future grid frequency variations, enabling the controller to provide forward-looking compensation capabilities. A virtual reference model is also constructed as a target trajectory generation module. This model generates ideal voltage and current operating trajectories based on the system's set performance targets and historical response behavior. This model serves as a reference path in the model prediction process, ensuring that the actual control output is aligned with the expected performance. A top-level optimization control objective is set for the top-level MPC. This function comprehensively considers multiple performance metrics, including output voltage deviation, output current deviation, virtual reference model parameter deviation, and power command adjustment. The output voltage deviation measures the difference between the drive's output voltage and the reference trajectory, reflecting the system's voltage regulation effectiveness. The output current deviation characterizes the system's load responsiveness, ensuring that the power supply tracks load changes and achieves dynamic current adjustment. The virtual reference model parameter deviation limits rapid changes in the reference trajectory, preventing control output oscillations caused by frequent reference model changes. The power command adjustment constrains the control system's response to power command changes, avoiding energy jitter caused by over-adjustment. By assigning weights to these four performance metrics, a multi-objective optimization function is established that balances system output accuracy, control stability, and energy efficiency. Based on the top-level optimization control objective and the predicted grid frequency, combined with the damping ratio constraint, the virtual reference model control coefficients and power command are optimized. The damping ratio, a compromise parameter between system response speed and stability, suppresses high-frequency oscillations in the control output and enhances the system's disturbance tolerance. During the optimization process, the rate of change of the reference model output curve matches the drive's physical response capability. Using a constrained optimization algorithm, the control gain coefficients and power commands of the virtual reference model are solved to obtain a set of optimized control coefficients and a corrected power command. Based on these optimized virtual reference model control coefficients and corrected power commands, a low-level model predictive controller is established to perform specific control action selection tasks. A finite control set is defined, consisting of N preset PWM waveform combinations. Each combination consists of a different duty cycle and switching strategy, corresponding to a different control instruction.For each control option, a recursive simulation of the system state is performed based on the state prediction model in the underlying controller. The system state evolution trajectory over the next several sampling periods under the current control command is predicted and compared moment-by-moment with the optimized virtual reference trajectory. By calculating the difference between the predicted output and the reference target, a set of state deviation vectors is formed, from which a deviation cost function is constructed. This function measures the accuracy of the control option's output in the current state and reflects its ability to fit the target path. To enhance the stability and physical sustainability of the control behavior, penalty terms are added to the deviation cost function, such as the additional cost of frequent switching of control actions or excessive switching rates. These additional terms improve the implementability and physical controllability of the control strategy. After completing the state prediction and cost evaluation for all control options, a set of performance evaluation values is obtained, which are sorted to select the control sequence with the best performance. In this sequence, the first control action is considered the optimal real-time control instruction at the current moment, with the advantages of minimizing the cost, output closest to the reference trajectory, and the highest control smoothness. This control action is immediately used to drive the PWM signal generator, which is converted into a specific duty cycle instruction and sent to the switching module of the LED driver to control the actual output behavior, thereby achieving instant adjustment and response of the system voltage and current.

[0030] In a specific embodiment, the execution step performs system state prediction and deviation cost calculation on each control option in the underlying model predictive controller to obtain a performance evaluation value for each control option may specifically include the following steps: Perform state mapping on the finite control set in the underlying model predictive controller to obtain the switch state vector; Perform recursive calculation of the system state for T sampling periods on the switch state vector to obtain the system state prediction sequence for the next T moments; The deviation between the system state prediction sequence and the reference trajectory of the virtual reference model is calculated to obtain the state tracking error index; A control switching penalty term is added to the state tracking error index to obtain a comprehensive performance index. Based on the comprehensive performance index, a feasible control sequence is searched to obtain the optimal control candidate solution. The stability of the optimal control candidate solution is checked twice to obtain the performance evaluation value of each control option.

[0031] Specifically, a state mapping is performed on the finite control set in the underlying model predictive controller. In the underlying model predictive controller, the finite control set consists of several preset PWM waveform combinations. Each control option represents a PWM duty cycle and switching frequency configuration within a specific sampling period. Since the PWM modulation process essentially controls the on-time and off-time ratio of power devices, each PWM control option corresponds to the switching state change of the driver's power transistors. To enable system predictive analysis of these options, the control combinations are mapped into standardized switch state vectors. Each PWM waveform is represented as a set of logical states, reflecting its on-off pattern within a sampling period. This vector structure serves as the control input for subsequent recursive calculations of the system state. The switch state vectors are sequentially applied to the system dynamic model, performing a recursive calculation of the system state prediction for T sampling periods. The initial system state is the actual current LED driver state, including key variables such as input capacitor voltage, output inductor current, and output capacitor voltage. Based on the PWM behavior represented by the switch state vectors, a recursive system of state-space equations is used to calculate the system state change for each control option at T future sampling points. This prediction process takes into account the actual operating characteristics of LED drivers, including nonlinear variations in electrical parameters, dynamic load response, input disturbances, and waveform response variations caused by switching frequency variations, ensuring the accuracy and operability of the prediction sequence. The system state prediction sequence reflects each control option's ability to regulate key variables such as output voltage and current within a future time window. The deviation between the system state prediction sequence and the reference trajectory of the virtual reference model is calculated to generate a system state tracking error metric. The virtual reference model provides idealized voltage and current variation paths that, based on power targets, system damping requirements, and frequency response characteristics, form an optimal operating state for approximation. The state tracking error metric calculates the difference between each state point in the prediction sequence and the corresponding reference point, and then weights these differences over time to form an error score that measures the control option's ability to approximate the target trajectory. Error terms are calculated separately for output voltage deviation, current deviation, and other dimensions, or combined using a weighted approach to form a comprehensive error vector, which reflects the overall control accuracy. A control switching penalty term is introduced within the state tracking error metric to account for the continuity of control behavior and hardware implementation costs. The control switching penalty term is designed to prevent drastic changes in control action during consecutive control cycles. Frequent PWM waveform switching, especially under high-frequency drive operation, can lead to increased electromagnetic interference (EMI) in power devices, decreased system stability, and even increased heat loss and lifespan. The difference between each candidate control option and the control action in the previous cycle is quantified, using the norm of the vector difference or the number of switching times as an indicator. This difference is then multiplied by a tuning weight to form a penalty component.This penalty term is then superimposed with the state error indicator to form the comprehensive performance index of the control option. Based on this comprehensive performance index, a search is conducted for feasible control sequences within the current sampling period. This search process employs full enumeration, heuristic search, or cost-sorting-based rapid screening to ensure that the optimal control candidate is found within the permitted computation time. This is the set of control actions with the lowest comprehensive performance index among all control options. This control action strikes a balance between state approximation capability, switching stability, and energy control, representing the most viable control strategy for the current system state. A secondary stability check is performed on the optimal control candidate. Based on the LED driver state-space model and the stability discriminant function, the control input of the candidate is substituted into the system model to determine whether the system's energy function derivative or response behavior satisfies stability constraints, such as the presence of a negative feedback loop, damping requirements, and current divergence avoidance. If the candidate still meets all safe operating conditions in the stability function evaluation, it is confirmed as the final feasible strategy. If not, the control option is reverted to the suboptimal solution in the performance evaluation sequence and stability verification is repeated until a control action that achieves both optimal performance and system stability is selected. All control options that have undergone performance calculation and stability verification are assigned a performance evaluation value. This value includes error evaluation results for control accuracy and response speed, and integrates the control switching stability index and stability criterion calculation results. Based on these evaluation values, the underlying model predictive controller accurately selects and dynamically updates the PWM control signal, ensuring that the LED driver maintains high-precision tracking of output voltage and current, stable control behavior, and efficient energy consumption control under complex operating conditions such as dynamic loads, grid fluctuations, and environmental disturbances.

[0032] The above describes the energy management optimization method based on the LED driver in the embodiment of the present invention. The following describes the energy management optimization system based on the LED driver in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, an energy management optimization system based on an LED driver includes: A parameter acquisition module is used to collect parameters and perform nonlinear mapping on the input and output ends of the LED driver to obtain a high-dimensional feature space of the LED driver's operating status; A parameter sharing module is used to input the parameters in the high-dimensional feature space into multiple local parameter mapping networks to perform parameter sharing and obtain a local energy feature vector; A feature processing module is used to input the local energy feature vector into a fully connected neural network for feature processing to obtain the PWM duty cycle adjustment, the switching frequency adjustment and the target PI controller parameter adjustment; Stability analysis module, used to perform stability analysis on PWM duty cycle adjustment, switching frequency adjustment and target PI controller parameter adjustment to obtain a stable parameter control strategy; The predictive control module is used to perform two-layer model predictive control based on the stable parameter control strategy and output real-time PWM control signals.

[0033] Through the collaborative efforts of these components, a high-dimensional feature space is constructed for the LED driver's operating state, enabling comprehensive perception of input, output, and internal power device parameters. This effectively captures the complex nonlinear characteristics of the system's operating state. Multiple local parameter mapping networks are used to implement parameter sharing, significantly improving parameter processing efficiency, reducing network redundancy, and enhancing the ability to extract features from similar physical quantities, enabling more accurate characterization of local energy characteristics. Feature processing through a fully connected neural network enables coordinated optimization of PWM duty cycle, switching frequency, and PI controller parameters, significantly improving system response speed and control accuracy compared to traditional single-parameter adjustment methods. A stability analysis mechanism, based on a stability discriminant function derived from a state-space model, ensures that all control parameter adjustments meet system stability constraints, effectively avoiding potential system instability during the optimization process. A two-layer model predictive control architecture combines long-term energy efficiency optimization with short-term real-time control. This architecture not only accounts for long-term factors such as grid frequency variations, but also rapidly responds to load changes, achieving full-time-domain optimal control for energy management.

[0034] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0035] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0036] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An energy management optimization method based on LED driver, characterized in that: include: Parameters of the input and output terminals of the LED driver are collected and nonlinearly mapped to obtain the high-dimensional feature space of the LED driver's operating status; Inputting the parameters in the high-dimensional feature space into multiple local parameter mapping networks to perform parameter sharing to obtain a local energy feature vector; Inputting the local energy feature vector into a fully connected neural network for feature processing to obtain a PWM duty cycle adjustment amount, a switching frequency adjustment amount, and a target PI controller parameter adjustment amount; Performing stability analysis on the PWM duty cycle adjustment amount, the switching frequency adjustment amount, and the target PI controller parameter adjustment amount to obtain a stable parameter control strategy; Based on the stable parameter control strategy, a two-layer model predictive control is executed to output a real-time PWM control signal.

2. The energy management optimization method based on LED driver according to claim 1, characterized in that: The parameter collection and nonlinear mapping of the input and output ends of the LED driver are performed to obtain a high-dimensional feature space of the operating state of the LED driver, including: The input voltage value, input current value and driver temperature value of the input end of the LED driver are collected to obtain the original parameter set of the input end; The output voltage value, output current value and LED lamp group temperature value of the output end of the LED driver are collected to obtain the original parameter set of the output end; Collect the temperature values of the power switch tube, inductor and rectifier bridge inside the LED driver to obtain the original parameter set of the internal power devices; Combining the input end original parameter set, the output end original parameter set, and the internal power device original parameter set to form an original operating parameter set of the LED driver; Performing nonlinear mapping on each parameter in the original operating parameter set to obtain a parameter characteristic vector, and performing differential value calculation and mapping on parameters of adjacent sampling points in the original operating parameter set to obtain a parameter dynamic characteristic vector; The parameter feature vector is combined with the parameter dynamic feature vector to form a high-dimensional feature space of the LED driver operation state.

3. The energy management optimization method based on LED driver according to claim 1, characterized in that: The step of inputting the parameters in the high-dimensional feature space into a plurality of local parameter mapping networks to perform parameter sharing and obtain a local energy feature vector comprises: According to the high-dimensional feature space, the input voltage-current-temperature parameter group, the output voltage-current-LED temperature parameter group, and the power device temperature parameter group are divided to obtain multiple groups of local correlation parameter sets; Normalizing the parameter values in each group of local correlation parameter sets to obtain normalized parameter values; Constructing multiple local parameter mapping networks, each of which includes an input layer, a hidden layer, and an output layer, and implementing a parameter sharing mechanism for the weight parameters of adjacent local parameter mapping networks, so that parameters related to the same physical quantity share weight values in different local parameter mapping networks, thereby obtaining a parameter-sharing local mapping network; Inputting the normalized parameter value into the parameter-sharing local mapping network, and obtaining a local feature output vector of each local parameter mapping network through forward calculation; The local feature output vectors are nonlinearly combined to obtain a local energy feature vector that characterizes the correlation characteristics between relevant parameters.

4. The energy management optimization method based on LED driver according to claim 1, characterized in that: The inputting the local energy feature vector into a fully connected neural network for feature processing to obtain a PWM duty cycle adjustment amount, a switching frequency adjustment amount, and a target PI controller parameter adjustment amount includes: Inputting the local energy feature vector into the input layer of a fully connected neural network to perform a linear combination operation to obtain an input feature representation; The input feature representation is input into a three-layer hidden layer structure for feature extraction and transformation to obtain a deep feature map; Performing a PWM duty cycle adjustment calculation based on the deep feature map to obtain a PWM duty cycle adjustment amount; Performing a switching frequency adjustment calculation based on the deep feature map to obtain a switching frequency adjustment amount; Performing a current loop PI controller parameter adjustment calculation based on the deep feature map to obtain a current loop PI controller parameter adjustment amount, and performing a voltage loop PI controller parameter adjustment calculation based on the deep feature map to obtain a voltage loop PI controller parameter adjustment amount; The current loop PI controller parameter adjustment amount and the voltage loop PI controller parameter adjustment amount are combined into a target PI controller parameter adjustment amount.

5. The energy management optimization method based on LED driver according to claim 1, characterized in that: The performing stability analysis on the PWM duty cycle adjustment amount, the switching frequency adjustment amount, and the target PI controller parameter adjustment amount to obtain a stable parameter control strategy includes: Constructing a state-space model based on the LED driver, wherein state variables of the state-space model include input capacitor voltage, output inductor current, and output capacitor voltage, and control variables include PWM duty cycle adjustment, switching frequency adjustment, and target PI controller parameter adjustment; Creating a system stability discriminant function according to the state space model, wherein the system stability discriminant function includes quadratic terms and cross terms of each state variable; Substituting the PWM duty cycle adjustment value, the switching frequency adjustment value, and the target PI controller parameter adjustment value into the state space model, calculating the derivative of the system stability discriminant function along the system trajectory, and obtaining a stability function; Performing a range analysis on the stability function to obtain stability constraints, and based on the stability constraints, performing a quadratic programming solution on the control parameters that do not meet the stability requirements to obtain control parameter adjustments that meet the stability requirements; The control parameter adjustment amount satisfying stability is combined with the PWM duty cycle adjustment amount, the switching frequency adjustment amount and the target PI controller parameter adjustment amount to obtain a stable parameter control strategy.

6. The energy management optimization method based on LED driver according to claim 1, characterized in that: The method of executing a two-layer model predictive control based on the stable parameter control strategy and outputting a real-time PWM control signal includes: Constructing a prediction model of a top-level model predictive controller in a two-level model predictive controller according to the stability parameter control strategy, wherein the prediction model includes an LED driver model, a grid frequency change prediction model, and a virtual reference model; Setting a top-level optimization control target for the top-level model predictive controller, wherein the top-level optimization control target includes an output voltage deviation term, an output current deviation term, a virtual reference model parameter deviation term, and a power command adjustment term; According to the top-level optimization control target and the grid frequency prediction value, combined with the damping ratio constraint condition, the virtual reference model control coefficient and the power command are optimized and calculated to obtain the optimized virtual reference model control coefficient and the corrected power command; Establishing a bottom-level model predictive controller based on the optimized virtual reference model control coefficients and the modified power command, and defining a finite control set of N preset PWM waveform combinations of the bottom-level model predictive controller; Performing system state prediction and deviation cost calculation on each control option in the underlying model predictive controller to obtain a performance evaluation value of each control option; The first control action of the optimal control sequence is selected based on the performance evaluation value, and a real-time PWM control signal of the LED driver is generated.

7. The energy management optimization method based on LED driver according to claim 6, characterized in that: The system state prediction and deviation cost calculation are performed on each control option in the underlying model predictive controller to obtain a performance evaluation value of each control option, including: Performing state mapping on a finite control set in the underlying model predictive controller to obtain a switch state vector; Performing a recursive calculation of the system state for T sampling periods on the switch state vector to obtain a system state prediction sequence for T moments in the future; Calculating the deviation between the system state prediction sequence and the reference trajectory of the virtual reference model to obtain a state tracking error index; Adding a control switching penalty term to the state tracking error index to obtain a comprehensive performance index, and searching for feasible control sequences based on the comprehensive performance index to obtain an optimal control candidate solution; A secondary stability check is performed on the optimal control candidate solution to obtain a performance evaluation value of each control option.

8. An energy management optimization system based on LED driver, characterized in that: For executing the energy management optimization method based on an LED driver according to any one of claims 1 to 7, the energy management optimization system based on an LED driver comprises: A parameter acquisition module is used to collect parameters and perform nonlinear mapping on the input and output ends of the LED driver to obtain a high-dimensional feature space of the LED driver's operating status; A parameter sharing module, configured to input the parameters in the high-dimensional feature space into a plurality of local parameter mapping networks to perform parameter sharing and obtain a local energy feature vector; a feature processing module, configured to input the local energy feature vector into a fully connected neural network for feature processing to obtain a PWM duty cycle adjustment, a switching frequency adjustment, and a target PI controller parameter adjustment; a stability analysis module, configured to perform stability analysis on the PWM duty cycle adjustment amount, the switching frequency adjustment amount, and the target PI controller parameter adjustment amount to obtain a stable parameter control strategy; The predictive control module is used to perform double-layer model predictive control based on the stable parameter control strategy and output a real-time PWM control signal.

Images