Power management control system and control method

Abstract

The invention discloses a power management control system and method. The system comprises a power module array, a load sensing and predicting unit, a dynamic power scheduling controller and a seamless switching execution and equalization circuit unit. According to the method, multi-dimensional sensing data and an external load plan are fused, and an artificial intelligence model is utilized to predict a short-term load demand; further, the optimal online combination and power distribution scheme of the power module is dynamically solved by adopting a rolling time domain optimization algorithm with the optimal overall energy efficiency of the system as the target; and finally, undisturbed switching of the power module is realized through voltage pre-synchronization, solid-state switch accurate control and load smooth transfer technologies. The problems that a high-power redundant power supply system is low in light load efficiency, slow in dynamic response and uneven in module service life are solved, intelligent power scheduling and hardware dynamic reconstruction based on prediction are achieved, and the energy efficiency, reliability and dynamic performance of the system are improved.

Description

Technical Field

[0001] This invention relates to the field of power management technology, and more specifically, to a power management control system and control method. Background Technology

[0002] High-power installed power supplies, as core power units of critical infrastructure, must be designed to simultaneously ensure high efficiency, high reliability, and rapid dynamic response. Traditional solutions often employ a fixed parallel architecture with N+X redundancy, i.e., configuring (N+M) power modules, where N meet the rated load and M serve as redundancy backups. This system suffers from the following inherent drawbacks: To ensure redundancy, multiple modules must operate in parallel even under light loads, with each module's operating point far from its high-efficiency zone, significantly reducing overall system efficiency and resulting in substantial energy waste; reactive response to load changes leads to delays and potential output voltage dips; module switching is typically based on simple current thresholds, easily causing frequent start-stops or oscillations; independent module operation prevents the system from anticipating load changes and pre-configuring power resources; current sharing control between modules is complex and easily introduces bus voltage disturbances during switching; fixed-operation modules bear the main load for extended periods, accelerating aging, while backup modules have low utilization rates, meaning overall system reliability is not optimally improved by redundancy. While existing technologies include solutions for starting and stopping modules by monitoring load current, these are essentially passive responses based on current values ​​and cannot solve the problems of predictability and smoothness. Existing research attempts to introduce simple efficiency curve optimization, but these are mostly limited to static allocation and are not deeply integrated with module health status, future load requirements, and seamless switching hardware mechanisms. Therefore, we need to design an intelligent power management solution that can proactively manage power resources, dynamically reconfigure hardware topology, and ensure high efficiency and high reliability across the entire load range. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a power management control system and control method to solve the problems existing in the background technology.

[0004] The above-mentioned technical objective of the present invention is achieved through the following technical solution: a power management and control system, comprising: The power module array consists of N+M parallel power modules with a unified digital communication interface. Each power module integrates an internal controller for reporting status parameters including real-time output power, efficiency, temperature and health status, and receiving external power allocation and switching commands. The load sensing and prediction unit is used to monitor and predict the total load demand of the system in real time, including multi-dimensional sensing sub-units and a load learning and prediction engine. The multi-dimensional sensing subunit is used to collect the system bus output current and voltage, the current of at least one key load branch, and the corresponding system heat sink temperature in real time. The load behavior learning and prediction engine has a built-in machine learning model based on recurrent neural networks or temporal convolutional networks. Its input is coupled to the output of the multi-dimensional sensing subunit, and it also receives load scheduling plan signals or equipment status change warning signals from the upper management system of the powered equipment. The load behavior learning and prediction engine is configured to extract and fuse features from the input time series data and output a predicted sequence of the total system load power within a future preset time window. A dynamic power scheduling controller, with its input connected to the load sensing and prediction unit and its output connected to the internal controller of each power module via a digital bus, is configured to perform the following operations: Based on the predicted sequence, the real-time efficiency curves reported by each power module, and the module health score calculated based on module runtime and thermal stress, the system aims to maximize overall system efficiency as the objective function, while simultaneously satisfying the safety margin of instantaneous load demand and maintaining an online redundancy of no less than N+X as constraints. A rolling time-domain optimization algorithm is used to dynamically solve for the optimal online combination of power modules and the target output power allocation scheme for each module. Then, based on the allocation scheme, a sequence of instructions for module start-up / stop, power ramp, and current sharing reference value adjustment, including precise timing, is generated. The seamless switching execution and equalization circuit unit specifically includes: a low-loss intelligent power switch array, wherein the positive output of each power module is connected to the common DC bus through a low on-resistance MOSFET solid-state switch directly driven by the dynamic power scheduling controller; and a synchronous drive circuit between modules to ensure that the rising and falling edges of the drive signal have microsecond-level synchronization accuracy when multiple solid-state switches operate in concert, so as to minimize bus voltage disturbances during switching transients.

[0005] Optionally, the load behavior learning and prediction engine specifically includes: The feature extraction and fusion subunit is used to process the raw signals from the multi-dimensional sensing subunit and the load plan information from external input through time alignment, normalization and sliding window processing to generate a multi-dimensional feature vector. At least one long short-term memory network layer is used to process the time-series dependencies of the multidimensional feature vectors and learn the periodicity, trend and event-triggered patterns of load changes. The output layer outputs the predicted load power values ​​for a specific future time series, along with the prediction confidence interval.

[0006] Optionally, the rolling time-domain optimization algorithm executed by the dynamic power scheduling controller aims to minimize the sum of the total power conversion losses of all online power modules while meeting the load demands of a series of consecutive future time points. The total power conversion loss is calculated based on the power allocated to each online power module at the corresponding time point and the efficiency value reported in real time at that power point; The constraints specifically include: a first constraint, that at each future time point, the sum of the allocated power of all online power modules must not be less than the sum of the predicted load power at that time point and a preset safety margin power; The second constraint is that at any future time point, the total number of healthy power modules in the on state must always be greater than or equal to the sum of the minimum number of primary modules N and the number of redundant modules X required by the system. The third constraint is that at each future time point, the allocated power of each online power module is limited to the high-efficiency load range defined by the real-time efficiency curve of that module, which is defined as the power range in which the efficiency value is not lower than a certain percentage of the peak efficiency. The fourth constraint is that the change in the allocated power of the same power module between adjacent time points shall not exceed the maximum allowable power change rate limit of that module.

[0007] Optionally, the working method of the seamless switching execution and equalization circuit unit specifically includes: Module commissioning process: Voltage pre-synchronization stage: Before closing the output solid-state switch of the module to be commissioned, the dynamic power scheduling controller instructs the internal controller of the module to precisely adjust its output voltage to be less than a first preset threshold value. Impact-free closing phase: Controlling the corresponding switches in the low-loss intelligent power switch array to close within microseconds to achieve grid connection of the module; Load smoothing transfer phase: By adjusting the digital current sharing bus reference value of this module and the original online module, part of the load is transferred to the newly added module with a preset power change slope until the target output power distribution scheme is achieved; Module cut-out process: Load smooth unloading stage: The digital current sharing bus reference value is adjusted in reverse to transfer the load of the module to be cut out to other online modules with a preset power change slope; Zero-current shutdown phase: When the output current of the module is detected to be lower than a second preset threshold, its output solid-state switch is controlled to turn off within microseconds. Enter standby mode: The main power conversion circuit of the module is turned off, while the power supply to its communication and control circuits is maintained, and the module enters a low-power standby mode.

[0008] Optionally, the module health score is calculated in real time using a weighted algorithm based on the cumulative running time, cumulative thermal cycle count, equivalent series resistance variation trend of key capacitors, and historical fault codes of each power module. The score result is used in the optimization decision of the dynamic power scheduling controller to prioritize the use of modules with high health and to reduce the output power of modules with low health or arrange them to enter maintenance standby state.

[0009] A power control method based on the above-mentioned power management and control system is characterized by comprising: step S1, real-time data fusion and load prediction step: collecting electrical parameters of the system bus and branches, combining them with external load planning signals, inputting them into a pre-trained load prediction model, and generating a high-precision load power prediction sequence within the future time window. Step S2, Dynamic decision-making steps based on multi-objective optimization: At the current decision moment, based on the predicted sequence, the efficiency mapping table and health score reported in real time by each power module, the primary optimization objective is to maximize the overall system efficiency, while the secondary objectives are to ensure N+X redundancy, limit the module power change rate, and balance the module lifespan. Rolling time-domain optimization calculation is performed to generate a forward-looking scheduling plan that includes the module switching state sequence and the power allocation sequence of each module. Step S3, Proactive Seamless Switching Execution Steps: When the scheduling plan indicates that a new module needs to be put into operation at a certain time in the future, the low-voltage pre-charging and voltage synchronization process of the module is started at a preset preheating time point before that time; at the planned operation time, its output solid-state switch is closed to allow it to be seamlessly connected to the bus; within a preset ramp time window, the load is smoothly transferred from the original module to the newly put into operation by adjusting the digital current sharing reference value; when the scheduling plan indicates that a module needs to be switched out at a certain time in the future, the reverse load transfer is performed within the ramp time window before that time, and the module is turned off with zero current at the planned switch-out time. Step S4, Closed-loop monitoring and adaptive learning steps: Real-time monitoring of the overall system efficiency, bus voltage ripple and module temperature rise after scheduling execution, comparing the actual data with the prediction and optimization targets; using the deviation data to fine-tune the parameters of the load prediction model online, and updating the model parameters accordingly.

[0010] Optionally, the low-voltage pre-charging and voltage synchronization process in step S3 is specifically implemented as follows: For the power module to be put into operation, its output capacitor is first pre-charged to near the bus voltage through a current-limiting resistor. Then, its main controller is activated and enters the fine voltage regulation mode. Through feedback control, the difference between its output voltage and the bus voltage is stabilized within the first preset threshold.

[0011] Optionally, the secondary objective of balancing module lifespan is achieved by introducing a penalty factor into the power allocation of each module in the rolling time-domain optimization calculation; the penalty factor is negatively correlated with the module's health score, so that modules with lower health scores are assigned a lower priority for increasing load during the optimization process, thereby achieving automatic load tilting towards healthy modules.

[0012] Optionally, it also includes a step of interacting with an external energy management system, the specific implementation process of which is as follows: The system receives external electricity price signals or total system power consumption limit instructions and uses these as additional constraints or optimization targets for the rolling time-domain optimization calculation in step S2, thereby achieving intelligent power scheduling based on economy or global power capping.

[0013] Optionally, in the closed-loop monitoring and adaptive learning steps, the module efficiency model is updated by accumulating the input and output electrical energy of each module and combining it with its real-time thermal imaging data to periodically calibrate the actual conversion loss curves of the module at different load points and temperatures.

[0014] In summary, the present invention has the following beneficial effects: 1. By proactively understanding load demand through load behavior learning and prediction engines, and executing rolling optimization centered on overall system efficiency via a dynamic power scheduling controller, this optimization algorithm can accurately calculate the optimal online combination of power modules that meets load and redundancy constraints at any future moment. This forces the power allocation of each online module to be constrained within its independently reported high-efficiency load range. For example, under light load conditions, the system can intelligently decide to deploy only the minimum number of modules and operate them at their high-efficiency points, while smoothly adding modules as needed when the load increases. This strategy can significantly improve the overall efficiency of the system under 20%-30% light load conditions compared to the traditional fixed multi-module current sharing mode, solving the problem of energy waste under light load conditions.

[0015] 2. By fusing multi-dimensional sensing subunits with external information, the system achieves advanced prediction of load changes. Based on this prediction, the system can pre-synchronize the voltage of the standby modules in the seamless switching execution and equalization circuit unit before the actual load surge, preparing them for grid connection. When a load change occurs, the system does not respond from zero, but instructs the pre-synchronized modules to achieve microsecond-level shockless closing through a low-loss intelligent power switch array, and immediately performs a smooth load transfer by adjusting the reference value of the digital current sharing bus. This proactive mode of prediction-preheating-execution transforms dynamic response from reactive remediation to proactive preparation, making the system's adjustment delay for step loads almost zero, and limiting bus voltage fluctuations to a very small range throughout the process, thus achieving high-quality dynamic response.

[0016] 3. Through a central dynamic power dispatch controller, discrete power modules, sensing networks, and load forecasting are integrated into an intelligent organism. The rolling time-domain optimization algorithm executed by the controller not only considers instantaneous efficiency but also coordinates the optimization of multiple objectives such as N+X redundancy reliability, module power change rate stability, and lifetime balancing based on health scores. In addition, through closed-loop monitoring and adaptive learning steps, the system continuously compares the predicted and actual values, and uses deviation data to fine-tune the AI ​​prediction model and efficiency model online, enabling the system to adapt to changes in load patterns and possess the ability to continuously self-optimize. This system-level intelligent collaboration optimizes traditional local and static control strategies.

[0017] 4. By introducing a module health scoring mechanism, which integrates multi-dimensional data such as cumulative operating time, thermal cycling, and degradation trends of key components, health is used as a key input in optimization decisions. By introducing a penalty factor negatively correlated with health, modules with low health are automatically reduced in power allocation, while modules with high health are prioritized for use. This mechanism intelligently tilts the workload towards modules in better condition and can also arrange for modules with low health to enter maintenance standby. As a result, wear leveling is achieved at the system level, avoiding premature failure at a single point, and elevating redundant configuration from simple backup to actively participating in lifespan management, thereby extending the overall mean time between failures (MTBF) of the system and improving work efficiency. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the system implementation process of the present invention; Figure 2 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0019] To make the objectives, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Several embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein.

[0020] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances. The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature.

[0021] In this invention, unless otherwise expressly specified and limited, "above" or "below" a second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of a second feature includes the first feature being directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" of a second feature includes the first feature being directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature. The terms "vertical," "horizontal," "left," "right," "above," "below," and similar expressions are for illustrative purposes only and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed or operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0022] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] This invention provides a power management and control system, such as... Figure 1 As shown, it includes: The power module array consists of N+M parallel power modules with a unified digital communication interface. Each power module integrates an internal controller for reporting status parameters including real-time output power, efficiency, temperature and health status, and receiving external power allocation and switching commands. The load sensing and prediction unit is used to monitor and predict the total load demand of the system in real time, including multi-dimensional sensing sub-units and a load learning and prediction engine. The multi-dimensional sensing subunit is used to collect the system bus output current and voltage, the current of at least one key load branch, and the corresponding system heat sink temperature in real time. The load behavior learning and prediction engine has a built-in machine learning model based on recurrent neural networks or temporal convolutional networks. Its input is coupled to the output of the multi-dimensional sensing subunit, and it also receives load scheduling plan signals or equipment status change warning signals from the upper management system of the powered equipment. The load behavior learning and prediction engine is configured to extract and fuse features from the input time series data and output a predicted sequence of the total system load power within a future preset time window. A dynamic power scheduling controller, with its input connected to the load sensing and prediction unit and its output connected to the internal controller of each power module via a digital bus, is configured to perform the following operations: Based on the predicted sequence, the real-time efficiency curves reported by each power module, and the module health score calculated based on module runtime and thermal stress, the system aims to maximize overall system efficiency as the objective function, while simultaneously satisfying the safety margin of instantaneous load demand and maintaining an online redundancy of no less than N+X as constraints. A rolling time-domain optimization algorithm is used to dynamically solve for the optimal online combination of power modules and the target output power allocation scheme for each module. Then, based on the allocation scheme, a sequence of instructions for module start-up / stop, power ramp, and current sharing reference value adjustment, including precise timing, is generated. The seamless switching execution and equalization circuit unit specifically includes: a low-loss intelligent power switch array, wherein the positive output of each power module is connected to the common DC bus through a low on-resistance MOSFET solid-state switch directly driven by the dynamic power scheduling controller; and a synchronous drive circuit between modules to ensure that the rising and falling edges of the drive signal have microsecond-level synchronization accuracy when multiple solid-state switches operate in concert, so as to minimize bus voltage disturbances during switching transients.

[0024] Furthermore, the load behavior learning and prediction engine specifically includes: The feature extraction and fusion subunit is used to process the raw signals from the multi-dimensional sensing subunit and the load plan information from external input through time alignment, normalization and sliding window processing to generate a multi-dimensional feature vector. At least one long short-term memory network layer is used to process the time-series dependencies of the multidimensional feature vectors and learn the periodicity, trend and event-triggered patterns of load changes. The output layer outputs the predicted load power values ​​for a specific future time series, along with the prediction confidence interval.

[0025] Furthermore, the rolling time-domain optimization algorithm executed by the dynamic power scheduling controller aims to minimize the sum of the total power conversion losses of all online power modules while meeting the load demands of a series of consecutive future time points. The total power conversion loss is calculated based on the power allocated to each online power module at the corresponding time point and the efficiency value reported in real time at that power point; The constraints specifically include: a first constraint, that at each future time point, the sum of the allocated power of all online power modules must not be less than the sum of the predicted load power at that time point and a preset safety margin power; The second constraint is that at any future time point, the total number of healthy power modules in the on state must always be greater than or equal to the sum of the minimum number of primary modules N and the number of redundant modules X required by the system. The third constraint is that at each future time point, the allocated power of each online power module is limited to the high-efficiency load range defined by the real-time efficiency curve of that module, which is defined as the power range in which the efficiency value is not lower than a certain percentage of the peak efficiency. The fourth constraint is that the change in the allocated power of the same power module between adjacent time points shall not exceed the maximum allowable power change rate limit of that module.

[0026] Furthermore, the working method of the seamless switching execution and equalization circuit unit specifically includes: Module commissioning process: Voltage pre-synchronization stage: Before closing the output solid-state switch of the module to be commissioned, the dynamic power scheduling controller instructs the internal controller of the module to precisely adjust its output voltage to be less than a first preset threshold value. Impact-free closing phase: Controlling the corresponding switches in the low-loss intelligent power switch array to close within microseconds to achieve grid connection of the module; Load smoothing transfer phase: By adjusting the digital current sharing bus reference value of this module and the original online module, part of the load is transferred to the newly added module with a preset power change slope until the target output power distribution scheme is achieved; Module cut-out process: Load smooth unloading stage: The digital current sharing bus reference value is adjusted in reverse to transfer the load of the module to be cut out to other online modules with a preset power change slope; Zero-current shutdown phase: When the output current of the module is detected to be lower than a second preset threshold, its output solid-state switch is controlled to turn off within microseconds. Enter standby mode: The main power conversion circuit of the module is turned off, while the power supply to its communication and control circuits is maintained, and the module enters a low-power standby mode.

[0027] Furthermore, the module health score is calculated in real time using a weighted algorithm based on the cumulative operating time, cumulative thermal cycle count, equivalent series resistance variation trend of key capacitors, and historical fault codes of each power module. The score result is used in the optimization decision of the dynamic power scheduling controller to prioritize the use of modules with high health and to reduce the output power of modules with low health or arrange them to enter maintenance standby state.

[0028] In a specific embodiment, the power module array consists of (N+M) electrically and communication-standardized power modules that can be connected in parallel. Each power module serves as an intelligent and controllable basic unit, and its specific structure is as follows: Power conversion and output stage: At the core of each power module is a high-efficiency DC / DC or AC / DC power conversion topology, such as a full-bridge LLC resonant converter, dual active bridge, or phase-shifted full-bridge. This converter transforms the input power, such as 400V DC or three-phase AC, into a stable low-voltage DC output, such as 48V or 12V. The output is then filtered by a low equivalent series resistance (ESR) capacitor before being led out to the module's positive and negative power output terminals. All modules' -OUT terminals are directly connected in parallel to the negative terminal of the common DC bus. The +OUT terminals of each module are not directly connected in parallel, but are first connected in series with a low-loss intelligent power switch before being connected to the positive terminal of the common DC bus. This design allows the output of each module to be independently connected to or disconnected from the bus. Each power module integrates a microcontroller (MCU) or digital signal processor (DSP) as its internal controller. This controller forms a closed loop with the various parts of the module through the following peripheral circuits: Analog signal acquisition circuit: Connects to a high-precision ADC to sample the input voltage / current and output voltage of the module in real time. , Temperatures of key components (such as switching transistor temperatures) Temperature of magnetic components ); PWM drive and regulation circuit: The controller generates a PWM signal, which is used by the driver chip to control the on and off of the main power switch, and the PWM duty cycle or frequency is adjusted by algorithms such as digital PID to stabilize the output voltage; Digital communication interface circuit: Used to implement a unified communication protocol, such as the physical layer and data link layer based on CAN bus, PMBus or Ethernet. This interface is used for bidirectional data exchange with the system's dynamic power dispatch controller; Switch drive circuit: Used to receive logic signals from the controller and drive the gate of the aforementioned MOSFET solid-state switch (Qᵢ) connected in series at the output terminal to realize soft switching of the module output; Current sharing bus interface: An analog or digital interface used to transmit load current information between multiple parallel modules to realize automatic current sharing between modules; The internal controller has embedded firmware, which mainly performs the following functions: Status monitoring and calculation module: Periodically (e.g., every millisecond) reads raw data from the ADC and calculates the real-time output power, expressed as... Real-time input power, Used for calculation based on the sampled values ​​from the input side; real-time conversion efficiency, expressed as... The constraint condition is when Rated threshold; health-related parameters: record running time, count the number of thermal cycles, and estimate the change in the equivalent series resistance of the output filter capacitor through online algorithms; Communication processing module: Parses commands received from the higher level through the digital communication interface, such as setting the output current reference value. Enable / disable commands and package and report the status data of this module according to a predetermined format and period (such as every 100 milliseconds); Power control and protection module: Executes voltage loop and current loop control algorithms to respond to valid signals. Then, it is compared with the machine's output current, and the actual output current is made to track the output current by adjusting the current sharing loop. Simultaneously, it implements multiple protection logics such as overvoltage, overcurrent, and overtemperature protection. Switch control module: Based on instructions, it controls the slow on / off (soft start) or off of the output solid-state switch Qᵢ, and can perform a voltage pre-synchronization process before the module is put into operation, finely adjusting the module. With system bus voltage The difference is within the millivolt level; As a whole, the power module array's workflow is characterized by the collaborative operation of all modules under the unified coordination of the dynamic power scheduling controller.

[0029] The core function of the load sensing and prediction unit is to realize the real-time monitoring and prediction of the total load demand of the system, which is specifically achieved by the multi-dimensional sensing sub-unit and the load behavior learning and prediction engine working together. The multi-dimensional sensing subunit provides the prediction engine with high-quality, multi-source, heterogeneous raw data streams. Its hardware configuration and data acquisition method are as follows: Bus total current and voltage sensing module: A high-precision, wideband current sensor based on the Hall effect is connected in series at the negative or positive terminal of the system's common DC bus; this sensor measures the system's total output current in real time. The sampling frequency is not less than 1kHz, and the measurement accuracy is better than ±0.5%. Bus voltage. The voltage is sampled through a highly stable resistor divider network, conditioned by an isolated operational amplifier, and then fed into an analog-to-digital converter (ADC). The bandwidth of this voltage sampling circuit needs to be higher than 10kHz to capture fast transients. Critical load branch current sensing module: To identify specific load patterns, such as the startup characteristics of compute-intensive or storage-intensive server clusters, a current sensor of the same accuracy is deployed on at least one representative load branch, for example, the power supply branch dedicated to GPU server clusters, to measure its branch current. The branch was selected because it accounts for a significant portion of the total load and its behavior pattern is representative. Ambient temperature sensing module: One or more digital temperature sensors are installed at the main heat dissipation airflow inlet or on the core heat sink of the system to measure the heat sink temperature. This temperature reflects the overall thermal environment of the system and is a key input for predicting load changes and assessing module heat dissipation conditions; External signal interface module: This is a digital communication interface, such as Ethernet, RS-485, or a specific system bus, used to receive load scheduling plan signals or equipment status change warning signals from the upper management system of the powered equipment, such as Data Center Infrastructure Management (DCIM) or Base Station Controller (BSC). The signals are structured digital information, such as data packets containing fields such as task ID, planned start timestamp, expected power consumption level, and equipment group shutdown warning. The load behavior learning and prediction engine is a software algorithm module that runs on the system's main controller. Its implementation is as follows: Data reception and buffer: The engine is equipped with multiple circular buffers, which are used to receive and temporarily store real-time time-series data streams from the above sensors and parsed external signals. The feature extraction and fusion subunit performs timestamp alignment and resampling. Since there may be slight deviations in the sampling times of different sensors, the engine first uses a unified system clock as a reference to align the timestamps of all input data streams. Subsequently, methods such as linear interpolation are used to resample all data onto a unified sequence with a fixed time interval Δt. Perform data cleaning and normalization on each time series data, for example... Outlier detection and removal are performed, followed by normalization, and the result is represented as follows: ,in, and The normalized boundary is determined based on historical data or sensor range; the external planning signal is encoded as a binary feature; Sliding window feature construction: At each prediction time t0, data from the past L time steps are extracted from each sequence to form a historical window; for each time point within the window, the normalized data and the encoded external signal features are concatenated to form a multi-dimensional feature vector, and the entire historical window is then constructed into a feature matrix of dimension [L,D]. , where D is the total dimension of the feature vector, which serves as the basis for inputting the prediction model; At the core of the engine is a pre-trained Long Short-Term Memory (LSTM) neural network model. Implementation details are as follows: The model input is the feature matrix generated in the previous step. The network structure of the model is as follows: it contains at least one LSTM layer; in a preferred embodiment, a two-layer stacked LSTM structure is used; the first LSTM layer has 128 neurons and is used to initially learn short-term temporal dependencies; the second LSTM layer has 64 neurons and is used to learn patterns and trends over longer periods; each LSTM unit contains an input gate, a forget gate, and an output gate, which selectively memorize and forget information through a gating mechanism, thereby effectively handling long-term dependencies of the load sequence. The model's output layer is the output of the last LSTM layer connected to a fully connected layer. The number of neurons in this layer equals the prediction time step T, and the activation function is a linear function. Its output is the predicted sequence of the total system load power at the next T time points, expressed as: ; In the output layer, the output layer can be modified to output the mean and variance of the predicted values ​​simultaneously, or techniques such as Monte Carlo Dropout can be used to perform multiple forward propagations during the inference stage and calculate the confidence interval of the predicted values ​​based on the result distribution. In the specific implementation process, this unit performs its work tasks according to a fixed control cycle: data acquisition is triggered by a system hardware interrupt or a timer. The instantaneous values ​​of the data synchronously sampled by each ADC in the multi-dimensional sensing subunit are collected; at the same time, the external signal interface module polls or interrupts to receive new messages from the upper-level system. All data is stamped with a unified high-precision timestamp and then sent to the buffer of the prediction engine. The feature extraction and fusion subunit is invoked to read the latest data from the buffer, perform timestamp alignment and resampling; apply calibration coefficients to each physical quantity sequence, convert it to standard units (A, V, °C), and perform data cleaning; load pre-stored normalization parameters and perform normalization calculations on all sequences; extract a historical data window of length L before the current time t0; convert external signals, such as a planned task starting in 30 seconds, into binary features (e.g., the feature is 1 if it will be active in the next 30 seconds, otherwise 0), and concatenate it with the sensor data features; thus generating the final... matrix; Will The matrix is ​​input into the loaded LSTM prediction model; the model performs forward propagation calculations, passing through the LSTM layer and the fully connected output layer in sequence. The T-dimensional vector output by the model is the load power prediction sequence for the next T seconds; the prediction engine encapsulates the generated sequence (and optional confidence intervals) into a data structure and delivers it to the dynamic power scheduling controller in real time through the internal message bus or shared memory, as the core input for its optimization decision-making; In the background tasks, to ensure the long-term adaptability of the entire model, before system deployment, historical load data, sensor data, and corresponding external event logs over a long period of time are used as the training set. The mean squared error (MSE) is used as the loss function, and the LSTM model is trained using backpropagation and time-truncation backpropagation (BPTT) algorithms until the prediction error of the model on the validation set converges. After the system is running, new running data (actual load values ​​as labels) are added to the dataset periodically (such as during daily off-peak periods) to incrementally train or fine-tune the model, so that the model can adapt to the slow evolution of load patterns.

[0030] The dynamic power dispatch controller is the central decision-making unit of the system, implemented using a high-performance embedded processor, a field-programmable gate array (FPGA), or a combination of both. Its software architecture includes the following core functional modules: The system status information aggregation module, acting as the central hub of the data bus, receives and integrates data from all other units in the system in real time, providing a complete input set for optimization decisions. Specifically, the aggregated information includes: The predicted data from the load sensing and prediction unit, representing the predicted sequence of total system load power within a future time window T, is expressed as follows: Where k represents the current decision time, and H represents the prediction / optimization time domain length, such as H=60, which represents the next 60 seconds; Real-time status data from the power module array is periodically received from each power module i (i=1,2,...,N+M) via a digital communication bus (such as CAN). The controller extracts and maintains the following key information for each module from this data: Current switch status, , representing the off state and the on state respectively; real-time efficiency characteristics, an efficiency function obtained by interpolation or fitting based on reported (power, efficiency) data points. This represents the conversion efficiency of module i when the output power is P. This function can be stored in the controller memory as a lookup table; implement health scoring. The value ranges from 0 to 100, and the score is provided for the power module array; other constraint parameters include the rated power of module i, the minimum / maximum allowable high-efficiency power point, and the maximum allowable power change rate, etc. The rolling time-domain optimization calculation module is the core of the controller algorithm. Its task is to periodically solve a multi-constraint optimization problem based on the aggregated information to determine the optimal scheduling plan in the future time domain. Specifically, it is implemented as follows: The optimization cycle is triggered periodically; this module is triggered at regular intervals with a fixed control cycle. Definition of decision variables: For each discrete time step t (t=k,k+1,...,k+H-1) and each power module i in the future optimization time domain, two decision variables are defined: switch state variables. , is a binary variable, where 0 indicates that module i is turned off by an instruction at time t, and 1 indicates that it is turned on by an instruction; output power variable , is a continuous variable, representing the power value output by module i at time t, in watts; The primary objective of constructing and calculating the objective function is to maximize the overall system efficiency, i.e., minimize the total energy conversion loss. The objective function J is specifically constructed as follows: Where J represents the total power conversion loss of the system in the optimized time domain H; This can be expressed as summing over H future time steps; This can be represented as summing over all (N+M) power modules; Let this be the switch state decision variable of module i at time t; Let the output power decision variable of module i at time t be represented; This is expressed as the efficiency of module i when the output power is P_i(t), which can be obtained by querying its efficiency function; This is expressed as the input power of module i at time t, calculated by back-calculating based on output power and efficiency. This is expressed as the power loss of module i at time t; This represents the actual time length corresponding to each time step, which is equal to the control period; the objective function aims to minimize the total heat loss caused by energy conversion in all power modules that are in the on state over a future period of time. For the optimization calculation of the above objective function, the following hard conditions apply: The first constraint is a power balance constraint, expressed as for all ,satisfy ,in This is expressed as the predicted load power at time t; This represents a preset safety margin of power to handle prediction errors and instantaneous overloads. This constraint ensures that the system's power supply capacity always meets the requirements. The second constraint is a redundancy reliability constraint, for all ,satisfy ,in This is represented as an indicator function, with a value of 1 when the condition inside the parentheses is true, and 0 otherwise. This is expressed as a health threshold (e.g., 70). This constraint ensures that at any given time, the number of modules that are active and healthy is not less than the minimum number of online modules required by the system (N+X). The third constraint is the efficient operating range constraint. For all i and t, if In this case, it is necessary to meet the following conditions. This constraint forces each online module to operate within its preset high-efficiency power range, avoiding inefficient operation; The fourth constraint is the power change rate constraint: for all i and t (t>k), it must satisfy... This constraint limits the variation of module power between two adjacent time steps, ensuring the smoothness of power regulation and reducing electrical stress on the module. The instruction sequence generation and distribution module optimizes the decision variable sequence output by the calculation module. This is transformed into specific instructions that can be directly executed by the power module and switching circuit; Instruction translation, for the decision result of the upcoming next time step (t=k), for each module i, according to... Generate the corresponding instructions, for example, if ,and Then, the module i startup command is generated, along with the target current value. , Represented as bus voltage; based on the switching state changes of all modules, a synchronous pulse sequence is generated to control the corresponding solid-state switch actions in the seamless switching execution circuit, and its trigger time is accurately calculated; Timing orchestration: To ensure seamless switching, this module performs fine-grained timing orchestration of instructions; for example, the module voltage pre-synchronization instruction will be issued Δt time before the "switch closing" instruction; the load transfer current reference value adjustment instruction will be issued gradually in a ramp form over multiple control cycles; Command issuance: The generated commands are issued in real time to the corresponding power module internal controller and FPGA-driven switch array through the corresponding interface (digital communication bus, GPIO, etc.).

[0031] In a specific embodiment, if the load sensing and prediction unit predicts that the load will jump from 2400W to 3200W in 60 seconds, the dynamic power scheduling controller, upon receiving the prediction sequence, initiates optimization in the current cycle; the input is the load for the next 60 seconds. The status of the four modules (assuming PM1 and PM2 are online, PM3 and PM4 are in standby, and all have a health level > 90) is determined by the solver. To meet future load demands and the 3+1 redundancy constraint, the optimal solution is to immediately activate PM3 and distribute the total load (2100W + margin) to PM1, PM2, and PM3 near the load ramp point, ensuring each module's power is approximately 700W (within the high-efficiency zone). The controller immediately generates commands: sends a start-up and pre-synchronization voltage command to PM3; calculates and schedules the synchronization pulse for closing PM3's output switch Q3 at 58 seconds (before the predicted load rise); and generates a sequence of ramp current reference values ​​for PM1, PM2, and PM3 from their current power to the target power. This allows the system to achieve the effect of advance adaptive adjustment.

[0032] The seamless switching execution and balancing circuit unit is a key hardware actuator connecting the power modules and the common DC bus. Its core function is to enable hot-swappable switching of the power modules, ensuring stable and shock-free bus voltage when changing the system power topology. It consists of the following two parts: The low-loss intelligent power switch array consists of (N+M) identical switching units, each corresponding to a power module, enabling independent control of its output path. The core switching device, at the heart of each switching unit, is a low on-resistance N-channel enhancement-mode power MOSFET, denoted as... (i=1 to N+M); The drain (D) of the MOSFET is connected to the positive output terminal (+OUT_i) of the corresponding power module i, and the source (S) is connected to the positive terminal (+BUS) of the common DC bus. A fast recovery diode is connected in anti-parallel between its source and drain as a supplement to the parasitic diode. The selection criterion for the MOSFET is that its on-state voltage drop at the maximum operating current is much smaller than the output voltage regulation of the power module. For example, for a 48V system, a MOSFET with Rds(on)<1mΩ is selected so that the on-state voltage drop at a current of 100A is less than 0.1V, thereby reducing the switching conduction loss to a negligible level. Drive and protection circuitry: Each MOSFET It is equipped with an independent gate drive circuit that receives switching control signals from the subsequent synchronous drive circuit; the drive circuit typically employs an isolated driver (such as one based on transformer or capacitor isolation) to provide sufficient gate drive current to achieve... Rapid turn-on and turn-off (rise / fall time on the order of tens of nanoseconds); the drive circuit integrates undervoltage protection, Miller clamping and other functions, and sets appropriate dead time logic to prevent accidental turn-on; Status monitoring feedback: The switching unit includes a high-side current sensing circuit, such as an isolation amplifier or a dedicated current sensing IC, for real-time monitoring of the current flowing through it. current The current signal is fed back to the dynamic power dispatch controller to accurately determine the current sharing status after the module is put into operation and the zero current condition before it is cut off. The core task of the inter-module synchronous drive circuit is to ensure the synchronization of multiple intelligent power switches ( to When coordinated action is required (e.g., two modules are put into operation at the same time to cope with a large step load), its switching action is highly synchronized in time to minimize transient disturbances in bus voltage and current caused by inconsistent switching timing. This includes: a high-precision timing generator, which is implemented by a field programmable gate array (FPGA) or a microcontroller with high-resolution PWM output, with a global, highly stable digital clock running inside the FPGA. Synchronization signal generation and distribution: The dynamic power dispatch controller sends the optimized switching action plan, including which switch needs to operate at which precise absolute time point or system clock cycle, to the FPGA. Based on this plan, the FPGA generates multiple fully synchronized switching control pulse signals under its internal clock synchronization. The microsecond-level synchronization accuracy means that the time deviation between the rising and falling edges of the pulse signal is strictly controlled to be less than a preset range. This is achieved through the unified clock network and layout of the I / O ports within the FPGA. Drive signal power amplification: The logic level synchronization pulse signal output by the FPGA, after being amplified and electrically isolated by the aforementioned independent gate drive circuits, is finally applied to each MOSFET. The gate; The seamless switching of execution and equalization circuit units, under the precise command of the dynamic power dispatch controller, executes the module's input and output process. The first preset threshold (voltage synchronization threshold) and the second preset threshold (zero current shutdown threshold) involved in the process are pre-set key parameters; In a specific embodiment, taking the input module i as an example, the specific implementation is as follows: Voltage pre-synchronization stage: Command issuance: The dynamic power dispatch controller sends a start and pre-synchronization command to the internal controller of the module to be put into operation through the communication bus. The command contains the target output voltage value. Module soft start and regulation: After receiving the instruction, the internal controller of module i starts its main power converter and enters a special voltage follower mode; in this mode, the voltage feedback loop reference value of the module is set to the current bus voltage (obtained through communication); the module finely regulates its output voltage through its internal negative feedback control; The internal controller of the module continuously compares the difference between the two. When the difference is less than the first preset threshold, the module replies the pre-synchronization completion signal to the dynamic power scheduling controller through the communication bus. This stage ensures that the voltage on the output capacitor of the module is almost equal to the bus voltage before it is connected, which solves the problem of capacitor surge charging current caused by voltage difference when the switch is closed in principle. Impact-free closing phase: Synchronous closing command. When the dynamic power dispatch controller receives the pre-synchronization completion signal and reaches the preset activation time, it sends a closing switch to the FPGA. Synchronization instructions; Switch synchronous conduction: The FPGA's timing generator sets the signal to active in its next global clock cycle; this signal, after being amplified by the driver circuit, enables the MOSFET... Rapid conduction; since voltage synchronization has been achieved through the aforementioned steps. The dV / dt and surge current at the moment of conduction are extremely small; state confirmation: the high-side current detection circuit of the switching unit monitors the current. Starting from 0, ascend to confirm that the module is connected to the bus; The load smoothing transfer phase, which lasts from when the switch is closed until the power stabilizes: Intervention of flow equalization reference value: In Simultaneously or shortly after closing, the dynamic power dispatch controller begins sending new current reference values ​​to the internal controllers of module i and other online modules via the digital current sharing bus. For the newly added module i, its Starting from 0, the power increases at a preset slope, i.e., a preset power change slope, such as an increase of 150A per second; for existing online modules, its Accordingly lowered; Digital current sharing control: The internal current sharing controller of each module controls the current sharing based on the received data. Based on the average current signal from the current sharing bus, each module quickly adjusts its output current. The output current of module i rises smoothly, while the output current of other modules falls smoothly, keeping the total load current constant. When the output current of all modules reaches the value corresponding to the target output power allocation scheme determined by the optimization decision, the load transfer is completed. Taking module j as an example, the specific implementation is as follows: During the load smoothing unloading phase, from issuing a cut-off command until the current approaches zero, the command is issued as follows: The dynamic power dispatch controller transmits the current reference value of the module j to be cut off via the digital current sharing bus. The current gradually decreases to 0 at a preset negative slope (e.g., a decrease of 150A per second); Current transfer: The internal current sharing controller of module j reduces its output current accordingly; At the same time, the current reference values ​​of other online modules increase accordingly to take over the transferred load, and the total load current of the system remains stable during this stage. During the zero-current turn-off phase, detection and switching actions are performed. Zero-current detection: The high-side current detection circuit of the dynamic power dispatch controller and / or switching unit continuously monitors the current reference value of module j. When the current reference value of module j is detected to be less than the second preset threshold, it is determined that a zero-current or near-zero-current state has been reached, and it is suitable for turn-off operation. Synchronous turn-off command: The controller immediately or in the next control cycle sends a turn-off switch to the FPGA. Synchronization instructions; Switch synchronization off: The FPGA invalidates the signal with microsecond-level synchronization precision. The drive circuit quickly... The gate voltage is pulled down. Turn-off; Since the current is already extremely small before turn-off, the voltage spikes and switching losses generated during the turn-off process are very small, achieving zero-current turn-off. The cut-out module enters standby mode, upon confirmation. After a reliable shutdown, the dynamic power scheduling controller sends a shutdown command to module j; entering low-power standby: the internal controller of module j shuts down its main power converter, keeping only communication and necessary monitoring circuits active, and enters a low-power standby state, waiting to be called again; The process of switching out or in is essentially the result of close collaboration between three entities: the dynamic power scheduling controller (decision and command), the power module internal controller (precise voltage and current regulation), and the seamless switching execution and balancing circuit (synchronization and isolation of the switch). Voltage pre-synchronization eliminates capacitive shocks caused by voltage differences, smooth load transfer eliminates sudden changes in inductive current, and the high-precision synchronous operation of the switch avoids oscillations and disturbances that may be caused by asynchronous operation of multiple modules.

[0033] A power control method based on the above-mentioned power management and control system is characterized by comprising: step S1, real-time data fusion and load prediction step: collecting electrical parameters of the system bus and branches, combining them with external load planning signals, inputting them into a pre-trained load prediction model, and generating a high-precision load power prediction sequence within the future time window. Step S2, Dynamic decision-making steps based on multi-objective optimization: At the current decision moment, based on the predicted sequence, the efficiency mapping table and health score reported in real time by each power module, the primary optimization objective is to maximize the overall system efficiency, while the secondary objectives are to ensure N+X redundancy, limit the module power change rate, and balance the module lifespan. Rolling time-domain optimization calculation is performed to generate a forward-looking scheduling plan that includes the module switching state sequence and the power allocation sequence of each module. Step S3, Proactive Seamless Switching Execution Steps: When the scheduling plan indicates that a new module needs to be put into operation at a certain time in the future, the low-voltage pre-charging and voltage synchronization process of the module is started at a preset preheating time point before that time; at the planned operation time, its output solid-state switch is closed to allow it to be seamlessly connected to the bus; within a preset ramp time window, the load is smoothly transferred from the original module to the newly put into operation by adjusting the digital current sharing reference value; when the scheduling plan indicates that a module needs to be switched out at a certain time in the future, the reverse load transfer is performed within the ramp time window before that time, and the module is turned off with zero current at the planned switch-out time. Step S4, Closed-loop monitoring and adaptive learning steps: Real-time monitoring of the overall system efficiency, bus voltage ripple and module temperature rise after scheduling execution, comparing the actual data with the prediction and optimization targets; using the deviation data to fine-tune the parameters of the load prediction model online, and updating the model parameters accordingly.

[0034] Optionally, the low-voltage pre-charging and voltage synchronization process in step S3 is specifically implemented as follows: For the power module to be put into operation, its output capacitor is first pre-charged to near the bus voltage through a current-limiting resistor. Then, its main controller is activated and enters the fine voltage regulation mode. Through feedback control, the difference between its output voltage and the bus voltage is stabilized within the first preset threshold.

[0035] Optionally, the secondary objective of balancing module lifespan is achieved by introducing a penalty factor into the power allocation of each module in the rolling time-domain optimization calculation; the penalty factor is negatively correlated with the module's health score, so that modules with lower health scores are assigned a lower priority for increasing load during the optimization process, thereby achieving automatic load tilting towards healthy modules.

[0036] Optionally, it also includes a step of interacting with an external energy management system, the specific implementation process of which is as follows: The system receives external electricity price signals or total system power consumption limit instructions and uses these as additional constraints or optimization targets for the rolling time-domain optimization calculation in step S2, thereby achieving intelligent power scheduling based on economy or global power capping.

[0037] Optionally, in the closed-loop monitoring and adaptive learning steps, the module efficiency model is updated by accumulating the input and output electrical energy of each module and combining it with its real-time thermal imaging data to periodically calibrate the actual conversion loss curves of the module at different load points and temperatures.

[0038] Specifically, the overall architecture and objectives of the efficiency model update mechanism enable the system to continuously optimize. A key component is maintaining an accurate power module efficiency model. This model characterizes the actual conversion efficiency of module i at different output power points P and different key component temperatures T. The initial model is established based on module factory data or laboratory calibration data. During long-term operation, due to factors such as component aging and environmental changes, the actual efficiency characteristics will drift. The established update mechanism collects actual operating data online and periodically calibrates the model to ensure that the efficiency information relied upon by the dynamic power scheduling controller remains highly accurate, thereby guaranteeing the effectiveness of optimization decisions. The implementation of this update mechanism relies on the fusion of three data sources: accurate accumulated energy data based on the coulomb counting method; real-time thermal imaging data; and module load operating point records, thereby achieving adaptive updates to the model.

[0039] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A power management and control system, characterized in that, include: The power module array consists of N+M parallel power modules with a unified digital communication interface. Each power module integrates an internal controller for reporting status parameters including real-time output power, efficiency, temperature and health status, and receiving external power allocation and switching commands. The load sensing and prediction unit is used to monitor and predict the total load demand of the system in real time, including multi-dimensional sensing sub-units and a load learning and prediction engine. The multi-dimensional sensing subunit is used to collect the system bus output current and voltage, the current of at least one key load branch, and the corresponding system heat sink temperature in real time. The load behavior learning and prediction engine has a built-in machine learning model based on recurrent neural networks or temporal convolutional networks. Its input is coupled to the output of the multi-dimensional sensing subunit, and it also receives load scheduling plan signals or equipment status change warning signals from the upper management system of the powered equipment. The load behavior learning and prediction engine is configured to extract and fuse features from the input time series data and output a predicted sequence of the total system load power within a future preset time window. A dynamic power scheduling controller, with its input connected to the load sensing and prediction unit and its output connected to the internal controller of each power module via a digital bus, is configured to perform the following operations: Based on the predicted sequence, the real-time efficiency curves reported by each power module, and the module health score calculated based on module runtime and thermal stress, the system aims to maximize overall system efficiency as the objective function, while simultaneously satisfying the safety margin of instantaneous load demand and maintaining an online redundancy of no less than N+X as constraints. A rolling time-domain optimization algorithm is used to dynamically solve for the optimal online combination of power modules and the target output power allocation scheme for each module. Then, based on the allocation scheme, a sequence of instructions for module start-up / stop, power ramp, and current sharing reference value adjustment, including precise timing, is generated. The seamless switching execution and equalization circuit unit specifically includes: a low-loss intelligent power switch array, wherein the positive output of each power module is connected to the common DC bus through a low on-resistance MOSFET solid-state switch directly driven by the dynamic power scheduling controller; and a synchronous drive circuit between modules to ensure that the rising and falling edges of the drive signal have microsecond-level synchronization accuracy when multiple solid-state switches operate in concert, so as to minimize bus voltage disturbances during switching transients.

2. The power management and control system according to claim 1, characterized in that, The load behavior learning and prediction engine specifically includes: The feature extraction and fusion subunit is used to process the raw signals from the multi-dimensional sensing subunit and the load plan information from external input through time alignment, normalization and sliding window processing to generate a multi-dimensional feature vector. At least one long short-term memory network layer is used to process the time-series dependencies of the multidimensional feature vectors and learn the periodicity, trend and event-triggered patterns of load changes. The output layer outputs the predicted load power values ​​for a specific future time series, along with the prediction confidence interval.

3. The power management and control system according to claim 1, characterized in that, The rolling time-domain optimization algorithm executed by the dynamic power scheduling controller aims to minimize the sum of the total power conversion losses of all online power modules while meeting the load demands of a series of consecutive future time points. The total power conversion loss is calculated based on the power allocated to each online power module at the corresponding time point and the efficiency value reported in real time at that power point; The constraints specifically include: a first constraint, that at each future time point, the sum of the allocated power of all online power modules must not be less than the sum of the predicted load power at that time point and a preset safety margin power; The second constraint is that at any future time point, the total number of healthy power modules in the on state must always be greater than or equal to the sum of the minimum number of primary modules N and the number of redundant modules X required by the system. The third constraint is that at each future time point, the allocated power of each online power module is limited to the high-efficiency load range defined by the real-time efficiency curve of that module, which is defined as the power range in which the efficiency value is not lower than a certain percentage of the peak efficiency. The fourth constraint is that the change in the allocated power of the same power module between adjacent time points shall not exceed the maximum allowable power change rate limit of that module.

4. A power management and control system according to claim 1, characterized in that, The working method of the seamless switching execution and equalization circuit unit specifically includes: Module commissioning process: Voltage pre-synchronization stage: Before closing the output solid-state switch of the module to be commissioned, the dynamic power scheduling controller instructs the internal controller of the module to precisely adjust its output voltage to be less than a first preset threshold value. Impact-free closing phase: Controlling the corresponding switches in the low-loss intelligent power switch array to close within microseconds to achieve grid connection of the module; Load smoothing transfer phase: By adjusting the digital current sharing bus reference value of this module and the original online module, part of the load is transferred to the newly added module with a preset power change slope until the target output power distribution scheme is achieved; Module cut-out process: Load smooth unloading stage: The digital current sharing bus reference value is adjusted in reverse to transfer the load of the module to be cut out to other online modules with a preset power change slope; Zero-current shutdown phase: When the output current of the module is detected to be lower than a second preset threshold, its output solid-state switch is controlled to turn off within microseconds. Enter standby mode: The main power conversion circuit of the module is turned off, while the power supply to its communication and control circuits is maintained, and the module enters a low-power standby mode.

5. A power management and control system according to claim 1 or 4, characterized in that, The module health score is calculated in real time using a weighted algorithm based on the cumulative operating time, cumulative thermal cycle count, equivalent series resistance variation trend of key capacitors, and historical fault codes of each power module. The scoring results are used in the optimization decision of the dynamic power scheduling controller to prioritize the activation of modules with high health and to reduce the output power of modules with low health or to arrange them to enter maintenance standby mode.

6. A power control method based on a power management and control system according to any one of claims 1-5, characterized in that, include: Step S1, Real-time Data Fusion and Load Forecasting: Collect electrical parameters of the system bus and branches, combine them with external load planning signals, and input them into the pre-trained load forecasting model to generate a high-precision load power forecasting sequence within the future time window. Step S2, Dynamic decision-making steps based on multi-objective optimization: At the current decision moment, based on the predicted sequence, the efficiency mapping table and health score reported in real time by each power module, the primary optimization objective is to maximize the overall system efficiency, while the secondary objectives are to ensure N+X redundancy, limit the module power change rate, and balance the module lifespan. Rolling time-domain optimization calculation is performed to generate a forward-looking scheduling plan that includes the module switching state sequence and the power allocation sequence of each module. Step S3, Proactive Seamless Switching Execution Steps: When the scheduling plan indicates that a new module needs to be put into operation at a certain time in the future, the low-voltage pre-charging and voltage synchronization process of the module is started at a preset preheating time point before that time; at the planned operation time, its output solid-state switch is closed to allow it to be seamlessly connected to the bus; within a preset ramp time window, the load is smoothly transferred from the original module to the newly put into operation by adjusting the digital current sharing reference value; when the scheduling plan indicates that a module needs to be switched out at a certain time in the future, the reverse load transfer is performed within the ramp time window before that time, and the module is turned off with zero current at the planned switch-out time. Step S4, Closed-loop monitoring and adaptive learning steps: Real-time monitoring of the overall system efficiency, bus voltage ripple and module temperature rise after scheduling execution, comparing the actual data with the prediction and optimization targets; using the deviation data to fine-tune the parameters of the load prediction model online, and updating the model parameters accordingly.

7. The power management control method according to claim 6, characterized in that, The specific implementation process of the low-voltage pre-charging and voltage synchronization process in step S3 is as follows: For the power module to be put into operation, its output capacitor is first pre-charged to near the bus voltage through a current-limiting resistor. Then, its main controller is activated and enters the fine voltage regulation mode. Through feedback control, the difference between its output voltage and the bus voltage is stabilized within the first preset threshold.

8. The power management control method according to claim 6, characterized in that, The secondary objective of balancing module lifespan is achieved by introducing a penalty factor into the power allocation of each module in the rolling time-domain optimization calculation. The penalty factor is negatively correlated with the module's health score, so that modules with lower health scores are assigned a lower priority for increasing load during the optimization process, thereby achieving automatic load tilting towards healthy modules.

9. The power management control method according to claim 6, characterized in that, It also includes the interaction steps with external energy management systems, and the specific implementation process is as follows: The system receives external electricity price signals or total system power consumption limit instructions and uses these as additional constraints or optimization targets for the rolling time-domain optimization calculation in step S2, thereby achieving intelligent power scheduling based on economy or global power capping.

10. A power management control method according to claim 6, characterized in that, In the closed-loop monitoring and adaptive learning steps, the module efficiency model is updated by accumulating the input and output energy of each module and combining it with its real-time thermal imaging data to periodically calibrate the actual conversion loss curves of the module at different load points and temperatures.

Images