A load-adaptive LED power supply control system and method
By employing a load-adaptive LED power supply control method, utilizing high-frequency synchronous sampling and lightweight neural networks to optimize grid operating conditions, and combining a total loss model and an adaptive PI controller, the efficiency fluctuations and operational discontinuities of LED driver power supplies under complex operating conditions are resolved, achieving energy efficiency optimization across all operating conditions and stable operation under extreme conditions.
Patent Information
- Application Number
- CN202611572181.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-12
- Publication Date
- 2026-11-10
AI Technical Summary
Existing LED driver power supplies exhibit large efficiency fluctuations when facing complex power grid conditions and diverse load demands, making it impossible to achieve system-level energy efficiency optimization across the entire operating range. Furthermore, they suffer from poor operational continuity under extreme conditions, which can easily lead to lighting flicker and shortened equipment lifespan.
A load-adaptive LED power supply control method is adopted. High-frequency synchronous sampling and fast Fourier transform are used to extract the power grid operating condition characteristics. Lightweight neural network is combined to classify the power grid quality and establish a total loss model. An adaptive PI controller and a lightweight reinforcement learning model are used to achieve dynamic optimization of bus voltage and operating mode, perform constant current control, and implement multi-level current reduction protection when the junction temperature of the power device meets the conditions.
It achieves optimal system-level energy efficiency across the entire operating range, eliminates mode jitter, ensures stable output current and lighting quality, and improves operational continuity and device lifespan under extreme conditions.
Smart Images

Figure CN123074243A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power electronic conversion and power supply control technology, specifically to an LED power supply control system and method based on load adaptation. Background Technology
[0002] LED driver power supplies are core components that determine the energy efficiency, lifespan, and reliability of lighting systems. LEDs themselves are low-voltage, high-current semiconductor devices, requiring dedicated constant-current / constant-voltage driver power supplies to convert grid power into DC output suitable for LED loads. Simultaneously, driver power supplies face two types of external constraints: firstly, the grid side conditions are complex and variable, including significant voltage fluctuations across regions and time periods, voltage sags and swells, harmonic distortion, surge impacts, and weak power quality environments in remote areas or industrial sites; secondly, the load side demands are diverse, with LED modules exhibiting a wide range of power and current requirements under different dimming levels, ambient temperatures, and series-parallel topologies.
[0003] In real-world power grids, non-ideal operating conditions are common, including voltage fluctuations, harmonic pollution, three-phase imbalance, voltage sags / surges, and frequency deviations. Since the operating point and control parameters of LED driver power supplies are fixed at the factory, high grid harmonic content exacerbates input current distortion and worsens the PFC correction effect. This not only leads to a deterioration in power factor (PF) and total harmonic distortion (THD), but also increases conduction losses and magnetic component losses. Furthermore, when voltage sags or surges occur, the system is prone to entering current-limiting protection or output instability states. This can cause minor issues like flickering and brightness fluctuations, or even trigger overvoltage / undervoltage protection, causing the entire lamp to shut off, severely impacting the continuity of lighting and the lifespan of electrical equipment. Meanwhile, traditional LED driver power supplies only reach peak efficiency within a narrow range near the rated input voltage. In the low-voltage input range, the input current amplitude increases, and the conduction losses of power devices and busbars rise significantly. In the high-voltage input range, device switching losses and electromagnetic interference intensify with increasing voltage stress. Existing power supplies use fixed switching frequencies and fixed voltage / current loop compensation coefficients, which cannot dynamically optimize the operating point based on real-time input voltage and load conditions. This results in a significant fluctuation in conversion efficiency across the entire wide input voltage range, characterized by "high efficiency in the middle and low efficiency at both ends," making it difficult to guarantee energy efficiency. Summary of the Invention
[0004] To address the problems of system efficiency collapse, mode switching jitter, and poor operational continuity caused by abrupt thermal protection in existing multi-stage power supplies under wide operating conditions, this application proposes a load-adaptive LED power supply control method to achieve optimal system-level global efficiency across the entire operating range, smooth and fluctuation-free output across the entire load range, and highly reliable continuous operation under extreme conditions.
[0005] To achieve the above objectives, this application adopts the following technical solution: a load-adaptive LED power supply control method, comprising the following operational steps: High-frequency synchronous sampling of voltage and current is performed on the input side. Harmonic spectrum is extracted using fast Fourier transform, effective voltage value and grid frequency are calculated, and a lightweight edge neural network is introduced to classify grid operating conditions in multiple ways and output grid quality feature vector. A total loss model is established, which includes the losses of the PFC pre-stage and the LLC stage. The bus voltage is used as the independent variable, and the real-time input voltage and output power are integrated to calculate the bus voltage that minimizes the total loss. The total loss model represents the relationship between the sum of the losses of the pre-stage power conversion circuit and the losses of the subsequent power conversion circuit and the change of the DC bus voltage. Based on real-time calculation of output power, system real-time conversion efficiency and current ripple rate by sampling on the output side, load levels are divided with rated power as the benchmark, and output load characteristic vector is generated. The load level and junction temperature are defined as state variables, and the pre-set operating modes are used as the set of selectable actions. A policy network model trained offline using a lightweight reinforcement learning model is used to determine the target operating mode of the subsequent power conversion circuit from multiple candidate operating modes. An adaptive PI controller with gain adjustment based on DC bus voltage and load rate is used, combined with output power feedforward, for constant current control; The calculation results of the total loss model and the ambient temperature of the device are reused to estimate the junction temperature of the power device, and multi-level current reduction protection is implemented when the junction temperature of the power device meets the preset overheating conditions.
[0006] Furthermore, the power grid quality feature vector includes five types of features: RMS input voltage, grid frequency, total harmonic distortion (THD), voltage sag / surge flag, and operating condition label.
[0007] Furthermore, the total loss model includes a front-stage loss sub-model and a rear-stage loss sub-model. The front-stage loss sub-model characterizes the relationship between the conduction loss and switching loss of the front-stage power conversion circuit and the DC bus voltage. The rear-stage loss sub-model characterizes the relationship between the gain deviation loss and frequency deviation loss of the rear-stage power conversion circuit and the DC bus voltage.
[0008] Furthermore, the calculation of the bus voltage that minimizes total loss includes: inputting the real-time input voltage, real-time output power, current DC bus voltage, and power device junction temperature into a pre-trained edge-lightweight large model, and outputting initial candidate values for the bus voltage; using the initial candidate values for the bus voltage as the initial values for iteration, and with minimizing total loss as the optimization objective, within the allowable constraint range of the bus voltage, performing numerical optimization iteration on the DC bus voltage using the total loss model, and determining the DC bus voltage that meets the preset convergence conditions as the optimal bus voltage setpoint.
[0009] Furthermore, load rating classification based on rated power includes: The current load rate is obtained by calculating the ratio of the actual output power to the rated output power, and the current load rate is mapped to the corresponding load level according to the preset threshold range. Load levels include: heavy load, medium load, and light load; The load feature vector includes five types of structured information: output power, system real-time conversion efficiency, load rate, current ripple rate, and load level.
[0010] Furthermore, determining the target operating mode of the subsequent power conversion circuit from multiple candidate operating modes includes: inputting system state variables containing load demand characteristics and device thermal states into a pre-trained strategy model, and outputting the expected loss evaluation value corresponding to each candidate operating mode; selecting the candidate operating mode with the optimal expected loss evaluation value as the initial target mode; maintaining the current operating mode of the subsequent power conversion circuit unchanged in response to the current load rate being within a preset hysteresis buffer; and determining the initial target mode as the target operating mode in response to the current load rate exceeding the preset hysteresis buffer.
[0011] Furthermore, constant current control includes: responding to an external dimming command, acquiring the target constant current reference value and the actual output current to calculate the current error; calculating a control command through an adaptive proportional-integral controller with feedforward compensation, which is used to drive the subsequent power conversion circuit; wherein, the proportional gain and integral gain of the adaptive proportional-integral controller are adaptively tuned online based on the current DC bus voltage and the current load rate; the adaptive proportional-integral controller with feedforward compensation is superimposed with a feedforward control term determined based on the target output power and the estimated output voltage.
[0012] Furthermore, the implementation of multi-stage current reduction protection includes: reusing the total loss calculated by the total loss model, combining the collected device ambient temperature and preset device thermal resistance parameters, and estimating the junction temperature of the power device through a first-order thermal resistance model; responding to the power device junction temperature being greater than the current reduction initiation temperature but less than the junction temperature safety limit, linearly reducing the upper limit of the output current according to the preset current reduction slope coefficient; and responding to the power device junction temperature being greater than or equal to the junction temperature safety limit, shutting off the output of the multi-stage power supply.
[0013] Furthermore, the implementation of multi-level current reduction protection also includes: implementing current limiting protection in response to the output current being less than the overcurrent trip threshold; and implementing shutdown protection in response to the output current being greater than or equal to the overcurrent trip threshold.
[0014] On the other hand, the present invention also provides a load-adaptive LED power supply control system, comprising: The front-end power conversion module is configured to receive input electrical energy and output DC bus voltage; The subsequent power conversion module is configured to receive the DC bus voltage and output electrical energy to the load; The digital control engine, connected to both the front-end and rear-end power conversion modules, is configured to execute a load-adaptive LED power control method.
[0015] Compared with the prior art, the beneficial effects achieved by the present invention are: (1) This invention solves the problem of efficiency mismatch between the front and rear stages under wide input and wide load conditions by constructing a total system loss model that includes the synergistic relationship between the losses of the front and rear stages and using the DC bus voltage as the independent variable for global optimization. This invention achieves optimal system-level energy efficiency across the entire operating range.
[0016] (2) By introducing a pre-trained strategy model and combining it with hysteresis buffer logic to make decisions on the working mode of the subsequent stage, this invention not only improves the accuracy and robustness of mode matching under complex working conditions, but also eliminates the mode jitter phenomenon at the load critical point from a mechanism perspective, ensuring the stability of the output current and the quality of the illumination light.
[0017] (3) This invention uses the calculation results of total system loss to estimate the junction temperature of power devices online, and combines it with a progressive current reduction and graded overcurrent protection strategy to achieve predictive flexible thermal management without increasing hardware costs. This effectively avoids the lighting interruption and thermal stress impact caused by traditional hard shutdown protection, and significantly improves the system's operational continuity and the service life of core devices under extreme conditions. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the LED power control system architecture provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the operation flow of the LED power supply control method provided in the embodiment of the present invention; Figure 3 This is a schematic diagram of the bus voltage optimization logic provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the intelligent selection logic for the working mode provided in an embodiment of the present invention. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.
[0020] Example 1: Combination Figure 1As shown, this embodiment provides a load-adaptive LED power supply control system. This system employs a front-stage adjustable bus PFC and a rear-stage resonant / synchronous rectified DC-DC converter, combined with a fully digital control and edge intelligence architecture. The front-stage uses PFC (Active Power Factor Correction) to correct the pulsating DC power after AC mains rectification into a stable DC bus voltage with high power factor and low harmonics, while also meeting grid harmonic requirements. The rear-stage topology is divided into LLC resonant topology and synchronous rectified Buck topology. The rear-stage converts the high-voltage DC bus into a low-voltage constant current output adapted to the LED load, while achieving electrical isolation or efficient voltage reduction. Edge intelligence senses the operating conditions in real time and calculates the globally optimal bus voltage and operating mode; the fully digital control is responsible for execution, converting AI decisions into precise PWM drive signals to control the power device operation, and forming a closed-loop control through sampling feedback.
[0021] In this embodiment, the system includes the following functional modules: Wide voltage input and PFC pre-amplifier module: As the core of the power input and pre-amplifier power conversion in the system of this invention, it is mainly used to receive the AC / DC wide voltage input from the front end, complete the rectification and conversion and active power factor correction, and convert the input power into a dynamically adjustable DC bus voltage.
[0022] Specifically, the wide-voltage input and PFC pre-amplifier module in this invention supports general AC mains input of 90V~305V and is also compatible with DC input scenarios of 350V~400V. Its output bus voltage no longer uses the traditional fixed 400V LED power supply; the upper and lower limits and target values of the output voltage are dynamically given by the upper-level bus optimization module. To achieve wide-range adjustable bus voltage, this module adopts a hybrid boost and buck-boost modulation topology. Addressing the limitation of traditional boost-type power factor correction PFCs that can only boost voltage, it can accurately set the bus voltage to the optimal operating point derived by the algorithm across the entire input voltage range. Simultaneously, the module integrates input voltage, current, and frequency sampling front-ends, directly providing raw sampling data to the grid condition sensing module without the need for additional sampling circuits. This provides a power execution foundation for the entire system's condition sensing and adaptive adjustment at the hardware level.
[0023] Bus voltage adaptive regulation module: It is used to receive the optimal target setpoint of the bus voltage from the edge large model and the online optimizer, generate the voltage reference command of the PFC controller, and drive the PFC controller to track the actual bus voltage to the target setpoint.
[0024] Specifically, the bus voltage adaptive adjustment module achieves decoupling and coordination between the front-end PFC circuit and the rear-end DC-DC circuit. By dynamically adjusting the bus voltage, it allows both power circuits to operate simultaneously within their respective high-efficiency operating ranges, serving as the core execution unit for achieving optimal global system efficiency. Under light load conditions in actual operation, this module actively reduces the bus voltage to decrease the stress and losses of the rear-end switches, while under heavy load conditions, it appropriately increases the bus voltage to ensure the power supply's load-carrying capacity.
[0025] MCU Digital Control Engine: In this invention, the engine runs a fully digital closed-loop control algorithm, including output constant current loop control, input voltage feedforward compensation, multi-mode scheduling logic, and fault protection logic. It can coordinate the working timing of all functional modules and ensure precise and coordinated operation of each module. At the hardware level, it integrates a high-speed analog-to-digital converter (ADC), a high-precision PWM wave generator, and various communication peripherals, which can meet all the hardware requirements for sampling, driving, and networking communication. At the same time, it reserves sufficient computing power resources to support local inference tasks of lightweight large models at the edge.
[0026] Specifically, in this invention, the MCU digital control engine replaces traditional fixed-parameter analog or digital compensation with a software-defined power approach. Software-defined power is a novel power architecture that decouples the control logic, functional configuration, and dynamic scheduling of a power system from fixed hardware. Instead, it uses software algorithms to perceive load and environmental changes in real time and dynamically adjust voltage, current, topology strategies, and energy efficiency modes. Based on this engine, all control parameters can be adjusted online in real time according to grid conditions, load status, and temperature environment, adapting to different application scenarios without modifying hardware circuitry, thus effectively improving the power supply's scenario adaptability.
[0027] Load detection and power identification module: It is used to collect voltage, current and temperature data of key power devices on the power output side in real time. Based on the sampled data, it calculates key parameters such as actual output power, overall conversion efficiency and output current ripple rate online. It also automatically identifies the current load level according to power and efficiency characteristics and divides it into three levels: heavy load, medium load and light load, and outputs them to the upstream decision module.
[0028] Adaptive control module: used to match the optimal power conversion operating mode for different load conditions, and achieve efficiency optimization across the entire load range.
[0029] Specifically, in the present invention, the adaptive control module performs smooth switching among three operating modes, namely the heavy-load synchronous rectification continuous mode, the medium-load phase modulation mode, and the light-load frequency hopping or burst mode, according to the load level and in combination with parameters such as the junction temperature of the power device and the current bus operating condition. A hysteresis threshold and an anti-shake logic are built in the switching process, which can avoid output fluctuation caused by frequent mode jumping and ensure stable LED output current without stroboscopic. In the heavy-load scenario, the synchronous rectification continuous operation mode is adopted, which reduces the conduction loss through the synchronous rectifier tube to ensure efficient operation under full power output; in the medium-load scenario, the phase modulation mode is adopted to adjust the output power, so as to effectively reduce the switching loss; in the light-load scenario, the intermittent frequency hopping or burst operation mode is adopted, which greatly reduces the standby, light-load and no-load losses, and specifically solves the problems of low light-load efficiency and high standby power consumption of the traditional LED power supply.
[0030] Multiple intelligent protection and thermal management module: configured to integrate full-dimensional fault protection and active thermal management mechanisms, so as to ensure safe and stable operation of the power supply under various extreme operating conditions.
[0031] Specifically, the multiple intelligent protection and thermal management module is built with basic fault protection mechanisms such as over-temperature current reduction, over-voltage protection, over-current protection, and short-circuit protection, which can quickly trigger a protection action when a fault occurs. Different from the hard protection mechanism of the traditional power supply that relies on a temperature threshold for triggering, this module estimates the junction temperature of the power device in real time based on the thermal model of the power device, and actively performs current reduction and current limiting actions before the junction temperature reaches the upper limit of the safety threshold, instead of hard turning off the power supply after a fault is triggered, thus realizing predictive protection. The active predictive protection mechanism in the present invention can avoid frequent start-stop and flicker disconnection of the power supply under operating conditions such as high temperature and overload, ensure continuous working capability under extreme operating conditions, and at the same time reduce the thermal stress impact on the power device, so as to prolong the service life of core components such as power semiconductors and electrolytic capacitors.
[0032] Dimming and communication interface module: which is an interactive interface for the system of the present invention to connect with an external control system and realize intelligent lighting networking, and is compatible with mainstream dimming and communication protocols in the current industry.
[0033] Specifically, in the present invention, the dimming and communication interface module supports 0-10V analog dimming and PWM hardware dimming, and also supports multiple digital dimming and communication protocols such as DALI, DMX, and private CAN bus, which can uniformly convert an externally input dimming command into a constant current reference value of the internal control loop, and send the same to the digital constant current closed loop to perform dimming control. Moreover, this module in the present invention also supports active reporting of power supply operating state data, which can realize networking and linkage control of multiple power supply devices. It ensures that the power supply can adapt to various application scenarios such as intelligent lighting and building automation, can realize smooth strobofree dimming in the full load range, can be seamlessly connected to intelligent lighting systems such as intelligent buildings, smart cities, and landscape lighting, and meets the requirements of centralized management and control and linkage scheduling of the lighting system.
[0034] Lightweight edge-based large model inference unit: This invention deploys a lightweight Transformer or neural network model that has been knowledge distilled and compressed in a local MCU chip. It adopts an offline training and online inference operation mode. Finally, the pre-trained lightweight edge-based large model takes multi-dimensional operating condition features such as input voltage, output power, current bus voltage, and device junction temperature as input, and outputs the overall efficiency prediction surface and power grid operating condition classification labels through online inference.
[0035] Specifically, in this invention, the unit replaces the traditional power-dependent lookup table method and offline parameter trial and error optimization method. The optimal operating point matching obtained based on multi-dimensional operating condition characteristics is faster and more accurate, and has stronger adaptability to complex operating conditions. At the same time, all inference calculations are completed at the local edge of the power supply, without relying on cloud network connection. A single power supply can independently complete intelligent adaptive adjustment, which can fully adapt to application scenarios without network coverage, such as industrial plants and outdoor street lights.
[0036] Example 2 In this embodiment, combined with Figures 2-4 As shown, the present invention provides an LED power supply control method based on load adaptation, comprising the following operational steps: Step S1: Perform high-frequency synchronous sampling of the input voltage and current, extract the harmonic spectrum, calculate the power quality parameters, perform multi-condition classification of the power grid, and output the power grid quality feature vector.
[0037] In this embodiment, the present invention uses a sampling frequency Input voltage on the input side Input current Perform high-frequency synchronous sampling to generate a length of The sliding window sampling sequence provides a unified time-series data source for subsequent frequency and time domain analysis. The input current within the window... Using Fast Fourier Transform (FFT) operations, the amplitude and phase information of each harmonic are decomposed. Further, the effective voltage value, real-time grid frequency, and total harmonic distortion (THD) are calculated using a sliding window root mean square (RMS) and zero-crossing detection algorithm. After parameter calculation, the normalized waveform feature vector is input into a TinyML 1D-CNN classification model for inference, outputting the corresponding grid operating condition classification label (normal / high harmonic / sag / boost / frequency shift) and classification confidence score. Finally, all extracted features are assembled into a grid quality feature vector. That is, including the effective value of the input voltage. Grid frequency Five types of characteristic quantities, namely total harmonic distortion (THD), voltage sag / boost flag, and operating condition label, are used to form structured power grid operating condition description data, thus completing the perception and feature extraction of the entire power grid.
[0038] Specifically, the input current within the window The Fast Fourier Transform (FFT) operation includes:
[0039] in, For the first The complex number calculation result corresponding to the root spectrum line includes both the amplitude and phase information of the corresponding subharmonic. For the first The values of each time-domain sampling point The value range is 0 to , The total number of sampling points within a single sliding window. The harmonic number is used when... The time corresponds to the fundamental component. It is a natural constant. The imaginary unit is used; this transformation can convert the current sampling sequence in the time domain to the frequency domain, separate the fundamental wave and each harmonic component, and obtain frequency domain data.
[0040] Based on the effective values of each harmonic obtained from the Fourier transform, the total harmonic distortion can be further calculated. :
[0041] in, For the first The effective value of the subharmonic current. To determine the highest harmonic order used in the statistical calculation, this embodiment uses the 50th order to cover the main harmonic components in the power grid. The effective value of the fundamental current is used. THD is ultimately quantified as a percentage to characterize the degree of distortion of the current waveform, serving as a core power quality indicator for assessing the degree of harmonic pollution in the power grid and measuring the effectiveness of PFC correction.
[0042] For the identification of transient operating conditions such as voltage sags and sags, the effective voltage value is calculated using the sliding window root mean square method:
[0043] in, This represents the effective voltage value calculated within the current sliding window. For the first The instantaneous values of voltage sampling points are N, where N is the total number of voltage sampling points within the sliding window. The effective voltage value sequence calculated by continuously updating the sliding window can accurately identify voltage dips and rises in the grid, capture the dynamic fluctuation process of voltage amplitude, and provide quantitative data for transient condition judgment.
[0044] Real-time power grid frequency is calculated using a zero-crossing detection algorithm. ,in The time interval between two adjacent voltage zero-crossing points is calculated by detecting the moment when the voltage waveform crosses zero, and then the real-time grid frequency is derived. This allows for the rapid identification of grid frequency deviation conditions and provides real-time frequency data for subsequent adaptive adjustment of control parameters.
[0045] Further, an original feature vector is constructed at the input of the classification process. This vector integrates all power quality features calculated at the front end, including one-dimensional sequence features such as the amplitude sequence of each harmonic of the current within the sliding window and the instantaneous voltage value sequence, as well as statistical features such as the effective voltage value, total harmonic distortion (THD), grid frequency, and voltage transient deviation, forming a unified one-dimensional original feature vector. Since the physical dimensions and numerical ranges of different features vary greatly (e.g., the effective voltage value is on the order of hundreds of volts, THD is a percentage value, and the frequency is on the order of fifty hertz), these dimensional differences can interfere with the accuracy of network feature extraction. Therefore, it is necessary to perform normalization processing on the original feature vector bit by bit, using min-max normalization to map all features to... Alternatively, the features can be mapped to a standard normal distribution using Z-score standardization, ultimately yielding a normalized input feature vector. A lightweight one-dimensional convolutional neural network is used to classify power grid operating conditions. The corresponding classification calculation is as follows:
[0046] in, For the final output of power grid operating condition classification labels, The input feature vector has been normalized. and These are the weight matrix and bias term of a one-dimensional CNN fully connected layer. The fully connected layer is responsible for mapping the extracted abstract features to the output dimensions of five operating conditions. After the linear transformation of the fully connected layer, the original scores of the five dimensions are output. Each score corresponds to one of the operating conditions: normal, high harmonic, temporary drop, temporary rise, and frequency shift. The scores can be positive or negative and do not have a probabilistic meaning. They only represent the degree of matching of the network to the corresponding category. The Softmax activation function is used to perform probabilistic transformation, which maps the output of the fully connected layer to the probability confidence scores corresponding to various working conditions. It maps the original score to a probability value between 0 and 1, and the sum of the probability values of all categories is 1. Each probability value is the classification confidence score of the current input belonging to the corresponding category. The index number is assigned to the operating condition category, and ultimately... The system calculates and outputs a classification label. This label is obtained by selecting the index corresponding to the category with the highest probability among five operating conditions, and using the operating condition name corresponding to that index as the final classification label. Simultaneously, the highest probability value represents the confidence level of the classification result. The control system can set a confidence threshold. When the highest confidence level is higher than the threshold, the classification result is accepted and the corresponding control strategy is matched. When the confidence level is lower than the threshold, the operating condition is considered uncertain, triggering a conservative control strategy, thereby improving the robustness of the system under complex operating conditions.
[0047] This invention transforms the non-intuitively quantifiable power grid quality status into a calculable and callable structured feature vector. This enables the control system to identify abnormal operating conditions such as high harmonics, sags, rises, and frequency shifts in advance, suppressing input current distortion and the deterioration of PFC correction effects. It effectively avoids output flickering and light outages caused by power grid disturbances, improves the continuity of power supply operation in complex power grid environments, and enhances core power quality indicators such as power factor and total harmonic distortion. Simultaneously, the introduced lightweight model can run locally in real-time on the edge MCU, without relying on cloud computing power, achieving accurate identification of various power grid operating conditions and meeting the deployment requirements of edge intelligence for power supply equipment.
[0048] Step S2: Establish a total loss model that includes the PFC pre-stage loss and the subsequent LLC loss. Using the bus voltage as the independent variable, integrate the real-time input voltage and output power to calculate the bus voltage that minimizes the total loss.
[0049] In this embodiment, addressing the problem in traditional fixed DC bus voltage schemes where the operating points of the preceding and following power circuits deviate from their optimal values and efficiency drops significantly over a wide input voltage range, this invention reads the power grid quality feature vector. Simultaneously, the output power data from the previous cycle is read as the input boundary conditions for the optimization calculation. The offline-trained lightweight large-scale model is invoked to infer the efficiency prediction surface under the current operating condition, outputting initial candidate values for the bus voltage to provide initial values for subsequent numerical optimization. Based on this, the optimization objective is to minimize the total system loss within the allowable constraint range of the bus voltage. Within the system, numerical optimization is performed using gradient descent or particle swarm optimization to calculate the optimal bus voltage setpoint. , obtained The output is sent to the bus voltage regulation module, which drives the PFC front-end to complete the closed-loop tracking control of the bus voltage, thereby achieving dynamic optimal adjustment of the bus voltage.
[0050] Specifically, the present invention establishes a total loss model that includes the PFC pre-stage loss and the subsequent LLC loss, including the PFC pre-stage conduction loss. :
[0051] in, The on-resistance of the PFC switch is... This refers to the effective value of the Boost-type inductor current, which in this invention is approximately equal to the effective value of the input current. This is the effective value of the input voltage. This is the DC bus voltage. The forward voltage drop of the boost diode. This refers to the power output of the power supply.
[0052] In this solution, PFC is used for the AC input scenario. This is the effective value of the input AC voltage, not the DC voltage. Therefore, the duty cycle formula for DC Boost cannot be directly applied. The instantaneous characteristics of sinusoidal AC need to be considered. The instantaneous value of the AC input voltage is... ,in It represents the peak value of the AC voltage; the conduction loss of the switching transistor needs to be averaged over a complete power frequency cycle. After engineering simplification, the average effect of the duty cycle will retain the value corresponding to the sinusoidal peak value. coefficients, ultimately forming The formula is in the form of a permutation of the on-time of the switching transistor within a power frequency cycle. Specifically, the first term of this formula corresponds to the on-time loss of the switching transistor, and its value varies with the DC bus voltage. The reason why the voltage decreases as the bus voltage increases is that an increase in the bus voltage reduces the duty cycle of the Boost circuit, shortens the conduction time of the switching transistor, and thus reduces the conduction loss; the second term corresponds to the conduction loss of the boost diode, which also decreases as the DC bus voltage increases. The reason why the average current of the diode decreases as the bus voltage increases is that the average current of the diode decreases as the bus voltage increases. Overall, the conduction loss of the PFC front stage decreases as the bus voltage increases.
[0053] PFC front-end switching losses :
[0054] in, The switching loss factor is determined by the rise time, fall time, and parasitic parameters of the switching transistor. This represents the peak current of the PFC switch. This refers to the switching operating frequency of the PFC circuit. The loss calculated in this invention varies with the DC bus voltage. The increase shows a linear upward trend, consistent with the aforementioned trend of increasing DC bus voltage. The conduction losses increase and decrease inversely, which means that the PFC front stage itself has an optimal operating bus voltage.
[0055] Losses of the subsequent LLC circuit :
[0056] in, The gain deviates from the loss coefficient. Frequency deviation loss coefficient, This refers to the primary and secondary turns ratio of an LLC transformer. The required output voltage for the LED load. This is the actual operating frequency of the LLC circuit. This is the inherent resonant frequency of the LLC circuit. When the DC bus voltage... Approximately n times the output voltage When the LLC circuit gain is close to 1 and the operating frequency is close to the resonant point, the circuit loss is minimal; when the DC bus voltage... When the value deviates from this value, the circuit gain deviates from the optimal value, the operating frequency deviates from the resonant frequency, and the loss increases accordingly. Therefore, the loss of the subsequent LLC stage will increase significantly as the bus voltage deviates from the resonant operating point.
[0057] DC bus voltage Using the single independent variable, the total system loss model is obtained by combining the preceding stage conduction loss, the preceding stage switching loss, and the subsequent stage LLC loss. :
[0058] Specifically, in this invention, numerical optimization is performed using gradient descent or particle swarm optimization to calculate the optimal bus voltage setpoint.
[0059] in, The learning rate is also known as the iteration step size, k is the number of iterations, and the partial derivative term is... The DC bus voltage is calculated using the aforementioned three types of loss formulas. Analytical differentiation yields the value representing the rate of change of total loss with voltage at the current bus voltage. During the iteration process, the algorithm continuously adjusts the bus voltage setpoint along the direction of decreasing loss. When the partial derivative is zero, the iteration reaches the minimum point of total loss, and the corresponding DC bus voltage at this point is the optimal setpoint. ,Right now .
[0060] To accelerate the optimization convergence speed and reduce the online computation load, this invention introduces a lightweight edge Transformer model to quickly predict the efficiency surface. The model expression is as follows:
[0061] in, This represents the overall system efficiency predicted by the model. This is a multidimensional feature embedding function used to embed the input voltage. Output power DC bus voltage Junction temperature of power devices Multidimensional features are converted into embedding vectors that the model can process. This refers to the junction temperature of the power device. This invention presents a lightweight Transformer model that has undergone knowledge distillation and compression. This model compresses the efficiency surface patterns inherent in massive offline training data into a lightweight model capable of online inference, providing initial values close to the optimal solution for gradient descent optimization. This can significantly reduce the dozens of iterations required by traditional technical solutions to just a few, improve the real-time performance of bus voltage optimization, and meet the high-speed response requirements of power control.
[0062] Step S3: Calculate the output power, system real-time conversion efficiency and current ripple rate based on the output side sampling, classify the load level based on the rated power, and output the load characteristic vector.
[0063] In this embodiment, traditional power supplies cannot accurately sense the load status and optimize operating strategies accordingly when the power and current demands of LED loads vary across a wide range of dimming depths, operating temperatures, and load topologies. This invention uses a high-precision sampling circuit to acquire the LED output voltage and constant current value, while simultaneously obtaining the system input power from the input-side sampling circuit. This provides the raw electrical data for subsequent calculations of all load characteristics. Based on the sampled parameters, the three core load characteristics—current actual output power, system real-time conversion efficiency, and output current ripple rate—are then calculated. After completing the characteristic calculations, the current load factor is further calculated. This is the ratio of actual output power to rated output power, and the load rate is mapped to the corresponding load level according to a preset threshold range. The threshold can be calibrated offline using a large model to adapt to LED load topologies of different specifications. Finally, all calculated feature quantities are assembled into a load feature vector. The vector contains the output power. Real-time system conversion efficiency Load rate Five types of structured information: current ripple rate, load level, etc.
[0064] Specifically, regarding output power With system real-time conversion efficiency Represented as:
[0065]
[0066] in, This is the DC voltage at the LED output terminal. This is the constant current value on the output side. The input power of the power supply system and the real-time conversion efficiency of the system. These reflect the power conversion capability of the power supply under the current operating conditions. In this invention, these two fundamental quantities are quantitative indicators for load identification, which are used for subsequent load level classification and can also be fed back to the bus optimization module for online correction of the loss model and verification of efficiency optimization results.
[0067] Current load rate Represented as: ,in This represents the rated output power of the power supply, indicating the proportion of the current actual output power to the rated power. In this embodiment, the corresponding classification rule is: when the load rate... When it is determined to be overloaded, When it is determined to be a medium load, The load is determined as light load. The three levels correspond to different power output ranges of the power supply. Heavy load corresponds to the power supply's near full power output, medium load corresponds to medium power output, and light load corresponds to low power and deep dimming. In this embodiment, the above-mentioned classification threshold is not a fixed value. It can be calibrated offline based on the efficiency characteristics of different LED load topologies using a large model, so that the classification threshold is more in line with the optimal efficiency inflection point of different application scenarios, further improving the scenario adaptability of the mode switching strategy.
[0068] Specifically, the current ripple rate in this invention Represented as:
[0069] in, and These represent the maximum and minimum values of the output current within a complete switching cycle. The ripple rate is the average output current during the switching cycle, quantified as a percentage to represent the degree of fluctuation in the output current. A lower ripple rate indicates a smoother constant current output characteristic, a more stable LED emission state, and a lower likelihood of flicker. In this embodiment, this indicator is used to evaluate the optical quality of the current dimming. It can also help determine the matching degree between the current load condition and the control mode. When the ripple rate exceeds a preset threshold, it can trigger an adjustment of the operating mode or a correction of the constant current loop parameters, ensuring flicker-free dimming across the entire load range.
[0070] Step S4: Define load level and junction temperature as state variables, and use the pre-set working mode as the set of optional actions. Use a policy network trained offline by a lightweight reinforcement learning (Q-learning / DQN) small model to select the working mode with the best loss under the current working condition online.
[0071] In this embodiment, the invention construct includes the current load level and the junction temperature of the power device. System state variables of the current bus voltage range This process integrates multi-dimensional working condition information into standardized state inputs that can be recognized by the reinforcement learning model. The state variables are then input into an offline-trained policy network, which outputs the expected loss for each selectable working mode action. Value, that is The quantitative characterization represents the long-term loss benefit corresponding to executing a certain operating mode under the current state. In this embodiment, the selectable operating modes include: heavy-load synchronous rectification continuous mode, medium-load phase modulation mode, and light-load frequency hopping burst or skip-cycle mode. Then, the optimal action that minimizes the expected loss is selected from all available actions. The system undergoes hysteresis-based logic verification to prevent frequent mode switching under boundary conditions. Finally, the selected optimal operating mode is mapped to specific low-level control parameters such as drive timing, duty cycle, and switching frequency, and then sent down to the subsequent DC-DC drive circuit for execution.
[0072] Specifically, the lightweight reinforcement learning (Q-learning / DQN) small model used in this invention is based on the reinforcement learning Q-value update rule, and the corresponding iterative formula is:
[0073] in, The current system state is represented by 'a', where 'a' represents the selected working mode action. The learning rate is used to control the magnitude of Q-value updates in each iteration, balancing training speed and convergence stability. For immediate reward, in this embodiment, the immediate reward is taken as the negative value of the total loss, i.e. This means that the smaller the loss, the greater the reward value obtained by the algorithm, thereby guiding the strategy to continuously optimize in the direction of minimizing loss. This is a discount factor used to weigh the current immediate rewards against future long-term rewards. The next state the system transitions to after performing the current action. This represents all possible actions for the next state. During training, the algorithm of this invention learns the optimal mode selection strategy for different operating conditions by continuously interacting with the simulation environment and iteratively updating the parameters of the Q-table or deep Q-network. After training, no further iterative updates are needed during the online running phase; only the current state needs to be directly input. By querying the Q value corresponding to each action and selecting the action with the largest Q value, the optimal working mode under the current working condition can be obtained. The online inference computation is low and can fully adapt to the computing power constraints of edge MCUs.
[0074] Specifically, to avoid anomalies caused by frequent mode fluctuations near the load threshold, this invention also introduces a mode switching determination mechanism with hysteresis. The corresponding segmented determination rule is: when the load rate... Optimal action Switch to heavy load mode; when the load rate Optimal action Switch to light load mode; maintain the original mode under other operating conditions. , These are the preset baseline thresholds for switching up and down modes, respectively. This refers to the hysteresis band width. In traditional single-threshold switching, if the load rate fluctuates slightly around the threshold due to sampling noise and load fluctuations, it can lead to frequent mode switching, causing audio noise, loss oscillations, and deterioration of output current ripple. By adding a hysteresis band in this invention, a gap is left between the rising and falling trigger thresholds for mode switching. Within the buffer zone, mode switching is only performed when the load rate continuously exceeds the buffer zone. When the load rate fluctuates within the buffer zone, the original operating mode remains unchanged, thus eliminating mode jitter from a mechanism perspective and ensuring the stability of the output current and the operational reliability of the power devices.
[0075] Meanwhile, for light-load frequency hopping burst modes, this invention precisely controls the conduction ratio during intermittent operation using the burst duty cycle formula, the corresponding calculation formula being:
[0076] in, This refers to the burst conduction time, which is the duration during which the power transistor operates normally in frequency hopping mode. Turn-off time, which is the interval during which the power transistor stops switching on; The circuit conversion efficiency in frequency hopping mode; The duty cycle in burst mode directly determines the actual energy output ratio of the power supply under light load conditions. Under light load conditions, the power supply operates in an intermittent burst manner, operating at high efficiency operating point parameters and outputting energy during the on-time period, and stopping the switching action of power devices during the off-time period, significantly reducing average switching losses and drive losses. This formula can accurately calculate the required duty cycle based on the current output power demand, minimizing the average loss under light load while meeting output power requirements and constant current accuracy, thus achieving a significant improvement in energy efficiency in the light load range.
[0077] Step S5: An adaptive PI controller with gain adjustment based on DC bus voltage and load rate is used, and output power feedforward is superimposed to achieve extremely low ripple constant current.
[0078] In this embodiment, to address the problems of flickering, insufficient constant current accuracy across the entire load range, and instability of the fixed compensation coefficient under wide voltage and load conditions in traditional LED power supply dimming processes, this invention receives external dimming commands and converts them to obtain the target constant current reference value. This serves as the target for closed-loop control. The actual output current is acquired in real-time via an output-side current sampling circuit. Calculate the current error between the target value and the actual value. This serves as the input to the closed-loop controller; then, an adaptive PI controller with feedforward compensation performs calculations to obtain the corresponding duty cycle or frequency control command. During the control process, the proportional and integral gains of the controller will be adjusted according to the current DC bus voltage. With load rate Online tuning is achieved to match the characteristics of the controlled object under the current operating conditions; finally, the calculated PWM wave or frequency command is sent to the subsequent DC-DC drive circuit to drive the power devices to operate, so that the output current continuously tracks the target current in a closed loop. It completes constant current and dimming control. It is also compatible with 0-10V analog dimming, PWM dimming and dimming of various digital protocols. By ensuring that the switching frequency of the subsequent stage is much higher than the visible light flicker frequency and controlling the output current ripple rate within the threshold, it achieves a smooth and flicker-free dimming effect across the entire load and dimming range.
[0079] Specifically, the adaptive PI controller with feedforward compensation performs the calculations to determine the corresponding duty cycle or frequency control command. for:
[0080]
[0081] in, The duty cycle or frequency control command output by the controller in the kth control cycle is a control quantity that directly acts on the subsequent drive circuit. , These are the proportional gain and integral gain of the controller, respectively. These are not fixed values and will be dynamically adjusted according to the current operating conditions. From the initial moment The sum of all historical current errors up to the current k-th cycle is the cumulative part of the integral term. The feedforward control term is obtained by dividing the target output power by the estimated output voltage. The function of the feedforward term is to provide control compensation in advance when there is a disturbance in the load or input, quickly offset the influence of the disturbance, greatly improve the dynamic response speed of the system, and reduce the current deviation under disturbance. This represents the current error in the current control cycle, reflecting the degree to which the output current deviates from the target. For the target constant current value, This is the measured output current.
[0082] Specifically, to address the issues of control instability and insufficient phase margin of fixed-gain PI controllers under wide input voltage and load range conditions, this embodiment also introduces a gain scheduling mechanism for online adaptive tuning of the PI parameters. The corresponding gain calculation formula is as follows:
[0083]
[0084] in, , The PI reference gain under the reference operating condition is a fundamental parameter of the controller. , This is a load rate scheduling function used to schedule tasks based on the current load rate. The proportional and integral gains are corrected to match the gain characteristics of the controlled object under different loads; , This is a bus voltage scheduling function used to adjust the current DC bus voltage. The PI gain is corrected to offset the control loop gain changes caused by bus voltage fluctuations. Through real-time correction of the dual scheduling functions of load rate and bus voltage, the controller can maintain sufficient phase margin and appropriate control bandwidth under all operating conditions with wide voltage input and wide load range. This avoids system oscillation and overshoot under light load and low voltage, and ensures response speed under heavy load and high voltage, eliminating the problem of instability of fixed parameter compensation under wide operating conditions.
[0085] Step S6: Synchronously collect key point temperature, voltage, and current in real time, estimate the junction temperature of power devices online based on the thermal resistance network model, and actively reduce current before the junction temperature of power devices approaches the safety limit; at the same time, configure multi-level protection for overvoltage / overcurrent / short circuit, and provide feedback and coordination of protection / thermal management results.
[0086] In this embodiment, the present invention synchronously collects the case temperature of the power device. Total real-time losses of the system Both are used as the basic input parameters for junction temperature estimation. The total loss data can directly reuse the calculation results of the loss model without repeated sampling and calculation. Then, the junction temperature of the power device is estimated online using a first-order thermal resistance model. This eliminates the need for a dedicated junction temperature sensor inside the device to obtain the core operating temperature of the power device. After estimating the junction temperature, over-temperature detection and adjustment are performed. If the junction temperature... Exceeding the downflow initiation temperature The upper limit of the output current will be gradually reduced according to the preset current reduction slope coefficient. The system replaces the traditional instantaneous hard shutdown with a gradual current reduction mechanism. When a serious electrical fault such as overvoltage, overcurrent, or short circuit occurs in the system, a graded shutdown protection mechanism is triggered, which ensures the safety of the devices while also taking into account the ability to resist false triggering. Finally, the information such as the protection action status and the current reduction operation status is written back to the preceding control module, triggering the bus voltage to be re-optimized to reduce the total system loss, or triggering the operation mode to switch to a low-loss, low-heat mode, forming a closed-loop synergy between protection and optimization.
[0087] Specifically, the junction temperature of the power device is estimated online using a first-order thermal resistance model. ,in, The case temperature of the power device or the ambient temperature can be directly acquired using a temperature sensor. The junction-to-case thermal resistance, which can be obtained from the device datasheet, is an inherent thermal characteristic parameter of power devices. This represents the current total system loss. In this embodiment, the first-order thermal resistance model is constructed based on the linear relationship between loss and temperature rise. It can predict the core temperature of the device online without the need for additional high-cost junction temperature sensors, identify overheating risks in advance, and provide quantitative temperature data for predictive thermal protection.
[0088] Specifically, when the system experiences serious electrical faults such as overvoltage, overcurrent, or short circuit, a tiered shutdown protection mechanism is triggered, including:
[0089] in, This is the rated output current of the power supply, which is the maximum output current value under normal operating conditions. The initiation temperature of the downflow is the temperature threshold that triggers active downflow; here... The current-dropping slope coefficient determines the rate at which the current decreases during the junction temperature rise. The junction temperature safety limit is the highest junction temperature that the device can withstand. This represents the upper limit of the output current after temperature limitation. In this invention, the piecewise curve achieves a smooth transition from full power output to linear current reduction and then to final shutdown. It replaces the hard shutdown protection of the traditional solution with gradual current reduction, which can reduce the thermal stress of the device when the temperature rise is too high, and can also maintain the continuity of lighting output to the maximum extent, avoid frequent start-stop interruptions, and ensure the lighting user experience.
[0090] In this embodiment, taking overcurrent protection as an example, the protection response time meets the following requirements. The corresponding protection action is divided into two levels: when the output current... Current limiting protection is implemented when the output current... The shutdown protection is executed at that time. Among them, The hardware delay for current sampling to fault identification, The delays in logic judgment and debouncing processing together determine the overall response speed of the protection. The overcurrent trip threshold is the dividing line between mild and severe overcurrent. This tiered action mechanism in the invention balances rapid protection with resistance to false triggering. For mild overcurrent conditions, current limiting maintains continuous system operation, while for severe overcurrent or short-circuit conditions, rapid shutdown is executed, reliably protecting power devices and LED loads from damage.
[0091] This invention is completely different from the traditional power supply protection mode that relies on hard threshold instantaneous shutdown. It achieves both full protection of the electrical safety of power devices and LED chips under extreme conditions such as high temperature, overvoltage, and short circuits, while maximizing the continuity of lighting output and avoiding frequent interruptions that affect the user experience. Simultaneously, by coordinating with preceding bus optimization and mode switching modules, it suppresses temperature rise at the source of losses, significantly reducing the thermal stress impact on devices and greatly extending the lifespan of core power supply components and LED chips, thus improving the overall system reliability and lighting user experience.
[0092] It should be understood that although this embodiment uses a proportional-integral controller as an example for explanation, in other embodiments, control algorithms with similar adaptive and feedforward capabilities, such as proportional-integral-derivative controllers, sliding mode controllers, or model predictive controllers, can also be used, as long as they can achieve stable constant current control under a wide range of operating conditions.
[0093] Example 3 To more intuitively demonstrate the performance advantages of the technical solution of this application in practical engineering applications, this embodiment provides a full-condition operation verification scenario in an LED intelligent lighting system. It should be understood that the following description is merely an illustrative application example and is not intended to limit the scope of protection of this application. Those skilled in the art can apply this load-adaptive LED power supply control method to other types of DC load power supply systems, such as server power supplies, battery chargers, or photovoltaic inverters, based on the teachings of this application, as long as they possess a multi-stage power conversion architecture and face energy efficiency and reliability challenges under a wide range of operating conditions.
[0094] In this application scenario, taking an outdoor LED street light driver power supply with a rated power of 200W as an example, the power supply adopts a two-stage architecture of front-end Boost PFC plus rear-end LLC resonant converter, and the digital control engine executes an adaptive control strategy.
[0095] First, this embodiment addresses the energy efficiency improvement issue under light load conditions at night. During the nighttime hours of urban road lighting, streetlights are typically dimmed to 10% or even lower load levels to save energy. In traditional control schemes, the power supply is often forced to operate in a continuously conducting mode with an extremely low duty cycle, or to adopt a simple skip-cycle mode, resulting in a sharp increase in switching losses. The conversion efficiency usually drops from over 95% at full load to below 70%, causing serious energy waste. In this embodiment, when the digital control engine detects that the load rate is below a preset threshold (e.g., 20%) and is in a stable dimming state, the pre-trained strategy model evaluates and selects an intermittent burst mode as the target operating mode. Specifically, the system accurately calculates the burst conduction duty cycle based on the ratio of real-time output power to rated output power, combined with the burst mode conversion efficiency. For example, at a 10% load rate, the power supply does not operate continuously with a 10% inefficient duty cycle, but rather outputs full-power pulses at the high-efficiency point near the resonant frequency for approximately 10.5% of the time, while completely shutting down the switching action of the power devices for the remaining nearly 90% of the time. This operating mechanism ensures that the power supply's equivalent operating point remains in the peak region of the efficiency curve under light load, thereby increasing the conversion efficiency to over 92% under 10% load. Compared with traditional solutions, this significantly reduces long-term electricity costs and carbon emissions.
[0096] Secondly, the efficiency and stability issues in complex power grid environments are addressed. In some older industrial areas or remote regions, the power grid voltage may experience significant fluctuations or harmonic pollution. If a traditional solution with a fixed 400V bus voltage is used, when the input voltage deviates from the rated value, the operating frequency of the downstream LLC circuit will be forced away from the resonant point to maintain voltage regulation. This leads to increased circulating current, increased core losses, and loss of soft-switching characteristics, potentially causing a 5%-10% drop in overall efficiency, accompanied by audible audio noise. In this embodiment, the digital control engine senses the power grid status in real time through the input sampling circuit and uses a total system loss model for global optimization. When a change in input voltage is detected, the efficiency prediction model quickly provides a new initial candidate value for the bus voltage. Subsequently, the numerical optimization algorithm precisely locks the target DC bus voltage that minimizes the total loss between the upstream and downstream stages within several control cycles. For example, when the input voltage drops from 220V to 180V, the system automatically lowers the bus voltage from 400V to approximately 360V, allowing the downstream LLC circuit to return to its resonant operating region with a gain close to 1. Actual test data shows that within the full input voltage range of 180V-265V, the overall efficiency fluctuation of this solution is less than 1.5%, and it maintains silent operation throughout, effectively solving the problems of energy loss and noise interference caused by poor grid adaptability.
[0097] Finally, regarding the issue of operational continuity under extreme high-temperature environments, outdoor lighting fixtures, when exposed to direct sunlight at midday in summer or installed in enclosed spaces with poor heat dissipation, can easily experience a rapid rise in the junction temperature of their internal power devices. Traditional power supplies typically employ hard threshold over-temperature protection, immediately cutting off the output once the temperature exceeds a safe limit (e.g., 110°C), and automatically restarting after the temperature drops. This frequent "hiccup"-like protection not only causes unsettling flickering in street lighting but also accelerates the aging and failure of electrolytic capacitors and solder joints due to repeated thermal shocks. In this embodiment, the system reuses the total system loss data calculated during bus optimization and combines it with a first-order thermal resistance model to estimate the junction temperature of power devices in real time, achieving sensorless predictive thermal management. When the estimated junction temperature exceeds the current-dropping initiation temperature (e.g., 95°C) but has not yet reached the safe upper limit, the system does not immediately shut down but linearly lowers the upper limit of the output current according to a preset current-dropping slope coefficient. For example, for every 1°C increase in junction temperature, the upper limit of the output current automatically decreases by 2% until the junction temperature stabilizes within the safe range. This gradual current reduction mechanism ensures that the lamps only experience a gradual decrease in brightness at high temperatures (the human eye is not sensitive to minute changes in brightness), rather than sudden shutdown. This guarantees basic road lighting functions and driving safety while avoiding subjecting the components to severe thermal stress cycles, significantly extending the lifespan of the power supply and LED modules. Only when an extreme anomaly causes the junction temperature to exceed the safe limit (such as 120°C) will the system execute the final shutdown protection as a last line of defense against catastrophic failures.
[0098] In summary, by applying a load-adaptive LED power supply control method to LED smart lighting scenarios, this invention successfully achieves high-efficiency operation across the entire load range, stable output with wide grid adaptability, and flexible safety protection under extreme conditions. These technical effects are not simply the sum of individual features, but rather stem from the deep synergy and data reuse among bus voltage optimization, intelligent mode switching, and predictive thermal management. This fully demonstrates the creative value and practical significance of the technical solution presented in this application in addressing real-world engineering challenges.
[0099] It should be noted that, in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.
[0100] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A load-adaptive LED power supply control method, wherein the LED power supply includes a front-end power conversion circuit and a back-end power conversion circuit, characterized in that, The following steps are included: High-frequency synchronous sampling of voltage and current is performed on the input side. Harmonic spectrum is extracted using fast Fourier transform, effective voltage value and grid frequency are calculated, and a lightweight edge neural network is introduced to classify grid operating conditions in multiple ways and output grid quality feature vector. A total loss model is established, which includes the losses of the PFC pre-stage and the LLC stage. The bus voltage is used as the independent variable, and the real-time input voltage and output power are integrated to calculate the bus voltage that minimizes the total loss. The total loss model represents the relationship between the sum of the losses of the pre-stage power conversion circuit and the losses of the subsequent power conversion circuit and the change of the DC bus voltage. Based on real-time calculation of output power, system real-time conversion efficiency and current ripple rate by sampling on the output side, load levels are divided with rated power as the benchmark, and output load characteristic vector is generated. The load level and junction temperature are defined as state variables, and the pre-set operating modes are used as the set of selectable actions. A policy network model trained offline using a lightweight reinforcement learning model is used to determine the target operating mode of the subsequent power conversion circuit from multiple candidate operating modes. An adaptive PI controller with gain adjustment based on DC bus voltage and load rate is used, combined with output power feedforward, for constant current control; The calculation results of the total loss model and the ambient temperature of the device are reused to estimate the junction temperature of the power device, and multi-level current reduction protection is implemented when the junction temperature of the power device meets the preset overheating conditions.
2. The LED power supply control method based on load adaptation according to claim 1, characterized in that, The power grid quality feature vector includes five types of features: RMS input voltage, power grid frequency, total harmonic distortion (THD), voltage sag / surge flag, and operating condition label.
3. The LED power supply control method based on load adaptation according to claim 2, characterized in that, The total loss model includes a front-stage loss sub-model and a back-stage loss sub-model. The front-stage loss sub-model characterizes the relationship between the conduction loss and switching loss of the front-stage power conversion circuit and the DC bus voltage. The back-stage loss sub-model characterizes the relationship between the gain deviation loss and frequency deviation loss of the back-stage power conversion circuit and the DC bus voltage.
4. The LED power supply control method based on load adaptation according to claim 3, characterized in that, The calculation of the bus voltage that minimizes total losses includes: Input the real-time input voltage, real-time output power, current DC bus voltage, and power device junction temperature into the pre-trained edge-lightweight large model, and output the initial candidate value of the bus voltage; Using the initial candidate value of the bus voltage as the initial value of the iteration and minimizing the total loss as the optimization objective, the DC bus voltage is numerically optimized and iterated using the total loss model within the allowable constraint range of the bus voltage. The DC bus voltage that meets the preset convergence condition is determined as the optimal bus voltage setting value.
5. The LED power supply control method based on load adaptation according to claim 4, characterized in that, Load rating classification based on rated power includes: The current load rate is obtained by calculating the ratio of the actual output power to the rated output power, and the current load rate is mapped to the corresponding load level according to the preset threshold range. Load levels include: heavy load, medium load, and light load; The load feature vector includes five types of structured information: output power, system real-time conversion efficiency, load rate, current ripple rate, and load level.
6. The LED power supply control method based on load adaptation according to claim 5, characterized in that, The target operating mode for the subsequent power conversion circuit is determined from multiple candidate operating modes, including: The system state variables, which include load demand characteristics and device thermal state, are input into the pre-trained strategy model, and the expected loss evaluation value corresponding to each candidate operating mode is output. The candidate working mode with the optimal expected loss assessment value is selected as the initial target mode; In response to the current load rate being within the preset hysteresis buffer range, the current operating mode of the subsequent power conversion circuit remains unchanged; In response to the current load rate exceeding the preset hysteresis buffer, the initial target mode is determined as the target operating mode.
7. The LED power supply control method based on load adaptation according to claim 6, characterized in that, Constant current control includes: In response to an external dimming command, the target constant current reference value and the actual output current are obtained to calculate the current error. Control commands are calculated by an adaptive proportional-integral controller with feedforward compensation, which are used to drive the subsequent power conversion circuit. Among them, the proportional gain and integral gain of the adaptive proportional-integral controller are adaptively tuned online based on the current DC bus voltage and the current load rate; The adaptive proportional-integral controller with feedforward compensation is superimposed with a feedforward control term determined based on the target output power and the estimated output voltage.
8. The LED power supply control method based on load adaptation according to claim 7, characterized in that, Implementing multi-level current reduction protection includes: The total loss calculated using the total loss model is combined with the collected device ambient temperature and preset device thermal resistance parameters to estimate the junction temperature of the power device using a first-order thermal resistance model. In response to the power device junction temperature being greater than the current-dropping initiation temperature but less than the junction temperature safety limit, the upper limit of the output current is linearly reduced according to the preset current-dropping slope coefficient. In response to the junction temperature of the power device being greater than or equal to the safe upper limit of the junction temperature, the output of the multi-stage power supply is shut off.
9. The LED power supply control method based on load adaptation according to claim 8, characterized in that, Implementing multi-level current reduction protection also includes: In response to the output current being less than the overcurrent trip threshold, current limiting protection is executed; In response to the output current being greater than or equal to the overcurrent trip threshold, shutdown protection is executed.
10. A load-adaptive LED power supply control system, characterized in that, include: The front-end power conversion module is configured to receive input electrical energy and output DC bus voltage; The subsequent power conversion module is configured to receive the DC bus voltage and output electrical energy to the load; A digital control engine, connected to both the front-end power conversion module and the rear-end power conversion module, is configured to execute the load-adaptive LED power control method according to any one of claims 1-9.
