A Smart Power Supply Regulation Method for Gaming Games Based on Load Prediction

Abstract

This invention discloses an intelligent power supply regulation method for e-sports PCs based on load prediction, belonging to the field of computer power management technology. It involves collecting real-time load data streams and game scene context data during the operation of an e-sports PC; decoupling the load features of the real-time load data streams; inputting the decoupled load features and game scene context data into a load prediction model based on a temporal convolutional network and attention mechanism to generate a predicted load stress index for each hardware component within a future time window; constructing a power distribution topology based on the motherboard power supply architecture; performing current distribution simulation on the power distribution topology; selecting the optimal current distribution path and phase configuration based on the predicted load stress index; generating a collaborative control instruction set; and issuing and executing the instruction. This invention achieves proactive power supply regulation through load prediction, effectively suppressing voltage transient drops and improving power supply efficiency.

Description

Technical Field

[0001] This invention relates to the field of computer power management technology, and in particular to a method for intelligent control of gaming power supply based on load prediction. Background Technology

[0002] In esports gaming scenarios, the load changes of computer systems are highly sudden and drastic. Rapid switching of game screens, simultaneous rendering of multiple units, bursts of complex physics calculations, and high-frequency peripheral operations can all cause the current demand of the central processing unit (CPU) and graphics processing unit (GPU) to jump from low load to full load within milliseconds. This drastic load change places extremely high demands on the transient response capability of the power supply system.

[0003] However, existing power regulation technologies have significant shortcomings in dealing with extreme load changes in esports scenarios. First, there is an inherent delay in regulation. From the occurrence of a load surge to the detection of a voltage drop and the execution of the regulation command, it typically takes tens to hundreds of microseconds. During this period, the voltage has already dropped significantly, causing a sharp drop in game frame rate or even system instability. Second, existing technologies have a relatively singular dimension in their perception of load, usually judging only based on the overall utilization of the CPU or GPU. They cannot distinguish the characteristics of different types of loads, such as graphics rendering, physics calculations, and peripheral interrupts. Different types of loads have significantly different transient response requirements for the power system, making it difficult to achieve precise matching using a uniform regulation strategy. In addition, the motherboard power supply network consists of multiple voltage regulation modules and multiple power supply phases, which are interconnected through power transmission paths. Existing technologies control each power supply unit independently, lacking systematic modeling and collaborative optimization of the overall power supply network. This makes it impossible to achieve joint power allocation of multiple phases and modules during load surges, resulting in excessive pressure on some power supply networks while other power resources are idle. This affects power supply efficiency and limits the system's ability to cope with load shocks. Summary of the Invention

[0004] The embodiments of the present invention provide a method for intelligent regulation of e-sports power supply based on load prediction, which aims to solve the problems of inherent delay, inability to distinguish load type and lack of power supply network collaborative optimization in the existing technology, resulting in voltage transient drops and low power supply efficiency.

[0005] To achieve the above objectives, this invention provides a method for intelligent power supply regulation based on load prediction in e-sports PC systems, comprising the following steps: The system collects real-time load data streams and game scene context data during the operation of a gaming PC. The real-time load data streams include CPU core utilization sequences, GPU core utilization sequences, video memory bandwidth utilization waveforms, and / or motherboard power supply phase current fluctuation trajectories. The game scene context data includes game frame rate, game scene tags, and / or peripheral input event frequency. By performing load feature decoupling processing on the real-time load data stream, the load features of graphics rendering, physical computing, and peripheral interrupts are separated. The decoupled load features and the game scene context data are input together into the load prediction model based on temporal convolutional networks and attention mechanisms to generate the predicted load pressure index of each hardware component within the future time window. Based on the motherboard power supply architecture and the physical connection of each hardware component, a power distribution topology is constructed. The nodes of the power distribution topology represent voltage regulation modules and the hardware components connected to them, and the edges represent power transmission paths and their dynamic impedance parameters. Perform current distribution simulation on the power distribution topology, calculate the overall power network loss and voltage transient response time under different current distribution schemes based on the predicted load pressure index, select the optimal current distribution path and phase configuration based on the simulation results, and generate a set of coordinated control instructions including voltage regulation module switching frequency adjustment, multi-phase current distribution ratio and load transient response compensation strategy. The coordinated control instruction set is sent to the power management microcontroller for execution.

[0006] Furthermore, the load characteristic decoupling processing of the real-time load data stream includes: Joint spectrum analysis is performed on the graphics processor core utilization sequence and memory bandwidth utilization waveform to extract graphics rendering cycle features and texture data stream features, and generate the graphics rendering load features. Perform multi-scale wavelet transform on the CPU core utilization sequence to extract rigid body dynamics calculation features and scene loading calculation features in physical calculations, and generate the physical calculation type load features. A time-series correlation analysis is performed on the frequency of peripheral input events and the changes in game frame rate to identify peripheral trigger-type load peaks and generate peripheral interrupt-type load characteristics.

[0007] Furthermore, the generation of the predicted load stress index for each hardware component within the future time window includes: The graphics rendering load characteristics, physics calculation load characteristics, peripheral interrupt load characteristics, and game scene context data are aligned along the time dimension to form a multi-channel time-series input matrix. A temporal convolutional network is used to extract multi-scale temporal features from the multi-channel temporal input matrix to capture load change patterns over different time spans. By using a multi-head attention mechanism to weightedly fuse the extracted temporal features, key time point features that are strongly correlated with future load peaks are obtained. The weighted and fused features are input into a fully connected prediction network, which outputs the predicted load pressure index of each hardware component within a future predetermined time window.

[0008] Furthermore, the construction of the power distribution topology includes: Scan the motherboard power management architecture to identify the voltage regulation module, the power supply phase of the central processing unit, the power supply phase of the graphics processing unit, and their respective hardware loads. Measure the dynamic impedance characteristics of each power transmission path under different current loads and establish an impedance-current function model. The voltage regulation module and hardware components are abstracted as nodes, and the node attributes include the predicted load pressure index; The power transmission path is abstracted as a directed edge, and the edge attributes include real-time current value, dynamic impedance parameters and the maximum current carried by the path. Based on the power supply connection relationship between the voltage regulation module and the hardware components, a weighted directed graph structure is established for the power distribution topology.

[0009] Furthermore, performing current distribution simulation on the power distribution topology includes: Based on the predicted load pressure index, identify hardware components that will experience a surge in current within a future time window as target nodes for regulation. In the power distribution topology diagram, all voltage regulation module nodes connected to the control target node are identified as candidate power supply nodes. Calculate the power loss and transient response delay along the power transmission path from each candidate power supply node to the control target node; The simulation allocates the current demand of the target node to multiple candidate power supply nodes in different proportions, generating various current allocation schemes. Calculate the total power loss and voltage drop recovery time of the power supply network under each current distribution scheme, and form a set of simulation results.

[0010] Furthermore, the calculation of power loss and transient response delay along the power transmission path from each candidate power supply node to the control target node includes: Calculate the resistive loss increment on the power transmission path based on the current current value, dynamic impedance parameters, and the current increment to be allocated. Based on the switching frequency of the voltage regulation module and the output capacitor parameters, calculate the transient response time of the output voltage when the current changes. Query historical power regulation records to obtain data on the actual voltage drop amplitude under similar current jump scenarios; The incremental resistive loss, transient response time, and historical voltage drop magnitude are weighted and integrated to generate a comprehensive power supply cost score.

[0011] Furthermore, the step of selecting the optimal current distribution path and phase configuration based on the simulation results includes: A multi-objective optimization function is defined, which simultaneously considers two objectives: minimizing the overall power supply network loss and minimizing the voltage transient response time. Pareto front analysis was performed on each scheme in the simulation results set to screen out the non-dominated solution set; In the non-dominated solution set, the optimization tendency is selected based on the current game scene type, which includes frame rate sensitive scenes and power consumption sensitive scenes. Taking into account the optimization tendency and the hardware execution overhead of phase switching, the scheme with the highest comprehensive score is selected as the optimal current allocation strategy.

[0012] Furthermore, the generation of the coordinated control instruction set, which includes voltage regulation module switching frequency adjustment, multi-phase current distribution ratio, and load transient response compensation strategy, includes: Based on the optimal current distribution strategy, determine the target switching frequency of each voltage regulation module; Based on the multi-phase current distribution ratio, the current output value of each power supply phase is reconfigured to achieve dynamic balance of load current among multiple phases. Based on the load transient response compensation strategy, the reference voltage or feedforward compensation parameters of the voltage regulation module are adjusted in advance to offset the voltage drop caused by the predicted load surge. The switching frequency adjustment instructions, current distribution instructions, and transient compensation parameters are encapsulated into a register configuration sequence that can be executed by the power management microcontroller.

[0013] Furthermore, the method also includes: The feedback learning process is optimized, and after executing the coordinated control instruction set, the actual voltage fluctuation curve, power supply efficiency data, and game frame rate stability data are continuously recorded. The actual data is compared with the expected values ​​in the simulation results to calculate the power supply regulation deviation rate; If the power supply regulation deviation rate exceeds the preset threshold, the online update of the load prediction model is triggered, and the weights of the temporal convolutional network and attention mechanism are fine-tuned using the newly collected data. The input characteristics, output commands, and actual effects of each adjustment are stored in the power optimization knowledge base for offline training and optimization of subsequent adjustment strategies.

[0014] Furthermore, the game scene context data also includes: the state of the game engine rendering pipeline, the predicted value of changes in the number of players in multiplayer online games, and the macro operation sequence of e-sports peripherals.

[0015] The above technical solution has the following technical effects: This invention addresses the problems of inherent delays, inability to distinguish load types, and lack of power network coordination in existing technologies, which lead to voltage transients and low power supply efficiency. These problems include: collecting real-time load data streams and game scene context data during the operation of a gaming PC; decoupling the load features from the real-time load data streams; inputting the decoupled load features and game scene context data into a load prediction model based on temporal convolutional networks and attention mechanisms to generate predicted load stress indices for each hardware component within future time windows; constructing a power distribution topology based on the motherboard power supply architecture; performing current distribution simulation on the power distribution topology; selecting the optimal current distribution path and phase configuration based on the predicted load stress indices; generating a collaborative control instruction set; and issuing and executing it. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an embodiment of the intelligent power supply control method for e-sports based on load prediction according to the present invention. Detailed Implementation

[0017] To further illustrate the various embodiments, the present invention provides accompanying drawings. These drawings are part of the disclosure of the present invention, primarily used to illustrate the embodiments and to explain the operating principles of the embodiments in conjunction with the relevant descriptions in the specification. With reference to these drawings, those skilled in the art should be able to understand other possible implementations and the advantages of the present invention. Components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0018] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0019] Figure 1 This is a flowchart illustrating an embodiment of the intelligent power supply control method for esports based on load prediction according to the present invention. Figure 1 As shown, the method of this embodiment includes the following steps: This system collects real-time load data streams and game scene context data from a gaming PC during operation. In one specific implementation, the real-time load data stream includes a CPU core utilization sequence, a GPU core utilization sequence, a video memory bandwidth utilization waveform, and a motherboard power supply phase current fluctuation trajectory. The CPU core utilization sequence is obtained through the operating system performance monitoring interface, with a sampling frequency of 1000 times per second to capture load fluctuations. The GPU core utilization sequence is obtained through the performance monitoring interface provided by the graphics driver, also with a sampling frequency of 1000 times per second. The video memory bandwidth utilization waveform is obtained through the performance counter of the graphics driver, recording the real-time bandwidth usage of video memory reads and writes. The motherboard power supply phase current fluctuation trajectory is obtained through the current sampling register inside the power management unit (Power Management Unit). Modern motherboard Power Management Units are typically integrated into the voltage regulation module controller chip. A read command is sent to the voltage regulation module controller via the system management bus to obtain the real-time current values ​​of each power supply phase. Specifically, the system management bus address and register offset are determined by the motherboard firmware. The power management microcontroller communicates with the voltage regulation module controller via the system management bus protocol to read the current sampling data.

[0020] Game scene context data includes game frame rate, game scene tags, and peripheral input event frequency. The game frame rate is obtained through the rendering statistics interface provided by the game engine. The game scene tags are obtained through the game engine's event callback mechanism; when the game scene changes, the game engine triggers a scene change event, and the system obtains the current scene tag by hooking into this event. The peripheral input event frequency is obtained through the event counter of the operating system's input subsystem, counting the number of input events from peripherals such as the mouse and keyboard at a frequency of 1000 times per second.

[0021] The real-time load data stream undergoes load feature decoupling processing to separate graphics rendering load features, physics computation load features, and peripheral interrupt load features. In one specific implementation, joint spectral analysis is performed on the graphics processor core utilization sequence and the memory bandwidth utilization waveform. A Fast Fourier Transform (FFT) is used to convert the time-domain signal into a frequency-domain signal, generating a power spectral density map. The graphics rendering cycle feature is manifested as significant peaks in the frequency components related to the game frame rate in the power spectrum; for example, when the game runs at 60 frames per second, obvious peaks appear at 60Hz and its harmonic frequencies. The texture data stream feature is manifested as a high-frequency fluctuation pattern in the memory bandwidth utilization waveform, corresponding to the loading and unloading process of texture data in the game scene. The extracted graphics rendering cycle features and texture data stream features are weighted and fused to generate graphics rendering load features.

[0022] A multi-scale wavelet transform is performed on the CPU core utilization sequence. Using the Daubechies wavelet basis, the utilization sequence is decomposed into detail coefficients and approximation coefficients at multiple scales. Rigid body dynamics computation features are characterized by sudden bursts in the high-frequency detail coefficients, corresponding to the load of collision detection and rigid body motion calculation in the physics engine. Scene loading computation features are characterized by a continuous rising pattern in the low-frequency approximation coefficients, corresponding to the load of loading large amounts of resources during game scene transitions. The extracted rigid body dynamics computation features and scene loading computation features are fused to generate physics-based load features.

[0023] A time-series correlation analysis was performed between the frequency of peripheral input events and changes in game frame rate. A sliding window cross-correlation method was used to calculate the correlation coefficient between the peripheral input event sequence and the game frame rate sequence at different time offsets. If the game frame rate drops significantly within a short period after a peripheral input event occurs, it is identified as a peripheral-triggered load peak. The temporal distribution and intensity of the identified peripheral-triggered load peaks were quantified to generate peripheral interruption load characteristics.

[0024] The decoupled load features and game scene context data are input into a load prediction model based on a temporal convolutional network and an attention mechanism to generate the predicted load pressure index for each hardware component within a future time window. In one specific implementation, the load prediction model adopts an architecture combining a temporal convolutional network and a multi-head attention mechanism. Graphics rendering load features, physics computation load features, peripheral interrupt load features, and game scene context data are aligned along the time dimension to form a multi-channel temporal input matrix. The time window length of the input matrix is ​​set to 100 milliseconds, with each time step corresponding to a 1-millisecond data sampling point.

[0025] The temporal convolutional network employs a multi-layered dilated convolutional structure, capturing load variation patterns over different time spans by setting different dilation rates. Shallow convolutional layers use kernels with a dilation rate of 1 to capture recent rapid load fluctuations; mid-level convolutional layers use kernels with dilation rates of 2 to 4 to capture load patterns over medium time spans; and deep convolutional layers use kernels with dilation rates of 8 to 16 to capture long-term load trends. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function.

[0026] A multi-head attention mechanism is used to weightedly fuse multi-scale temporal features extracted by a temporal convolutional network. Eight attention heads are set up, each independently calculating the linear transformation of the query, key, and value, and calculating attention weights. These attention weights reflect the importance of features at different time points to the prediction target. Through the fusion of multi-head attention, the model can strengthen key time-point features that are strongly correlated with future load peaks.

[0027] The weighted and fused features are input into a fully connected prediction network. This network consists of three fully connected layers: the first layer maps the feature dimension to 256 dimensions, the second to 128 dimensions, and the third to the output dimension. The output dimension corresponds to the prediction load stress index of each hardware component, including the CPU core prediction load stress index, GPU core prediction load stress index, and memory prediction load stress index. The prediction time window is set from 50 milliseconds to 500 milliseconds, with the specific value dynamically adjusted according to the game scene type. A shorter prediction window is used for frame rate-sensitive scenes to improve prediction accuracy, while a longer prediction window can be used for power-sensitive scenes to allow for earlier power scheduling.

[0028] Based on the motherboard power supply architecture and the physical connections of various hardware components, a power distribution topology is constructed. In one specific implementation, the motherboard power management architecture is first scanned to identify the voltage regulation module, the CPU power supply phase, the GPU power supply phase, and their corresponding hardware loads. Motherboard power management architecture information can be obtained by reading the power management descriptor table in the motherboard firmware or by querying the power management microcontroller via the system management bus.

[0029] The dynamic impedance characteristics of each power transmission path under different current loads were measured, and an impedance-current function model was established. Specifically, for each power transmission path, different magnitudes of step current loads were applied under laboratory conditions, the voltage drop waveforms were recorded, and the dynamic impedance of the path was calculated using Ohm's law. The dynamic impedance exhibits a non-linear relationship with the current load; typically, the impedance increases due to temperature rise under high current loads. An impedance-current function model was established, expressing the dynamic impedance as a function of the current load: Z(I) = Z0 + kI, Where Z0 is the basic impedance and k is the temperature rise coefficient.

[0030] Voltage regulation modules and hardware components are abstracted as nodes, with node attributes including the predicted load stress index. Voltage regulation module node attributes include current switching frequency, output capacitor parameters, and maximum output current capability; hardware component node attributes include the predicted load stress index, current current consumption, and voltage tolerance range. Power transmission paths are abstracted as directed edges, with edge attributes including real-time current value, dynamic impedance parameters, and the maximum current carried by the path. The direction of the edge points from the voltage regulation module to the hardware component, indicating the direction of current flow. For multi-phase power supply systems, multiple voltage regulation module nodes can point to the same hardware component node, indicating that multiple power supply phases jointly supply power to the same load. Based on the power supply connection relationships between voltage regulation modules and hardware components, a weighted directed graph structure is established for power distribution topology.

[0031] Perform current distribution simulation on the power distribution topology, calculate the overall power network loss and voltage transient response time under different current distribution schemes based on the predicted load pressure index, select the optimal current distribution path and phase configuration based on the simulation results, and generate a set of coordinated control instructions including voltage regulation module switching frequency adjustment, multi-phase current distribution ratio and load transient response compensation strategy.

[0032] In one specific implementation, the current distribution simulation process is as follows: Based on the predicted load stress index, hardware components that will experience a current surge within a future time window are identified as target nodes for regulation. For example, if the predicted load stress index of a GPU core jumps from its current value of 0.3 to 0.9 in 50 milliseconds, corresponding to an increase in current demand from 30 amps to 90 amps, then this GPU core is identified as a target node for regulation. In the power distribution topology, all voltage regulation module nodes connected to the target node are identified as candidate power supply nodes. For multi-phase power supply systems, multiple voltage regulation modules typically power the GPU core together; these modules are all considered candidate power supply nodes.

[0033] Calculate the power loss and transient response delay along the power transmission path from each candidate power supply node to the target control node. The power loss calculation formula is P. loss = I 2 ·Z(I), where I is the current along the path and Z(I) is the dynamic impedance. The transient response delay is calculated using the formula T. response = (ΔI·L) / ΔV max Where ΔI is the change in current, L is the output inductance value, and ΔV max The maximum allowable voltage drop is defined. The simulation distributes the current demand of the target node to multiple candidate power supply nodes in different proportions, generating various current allocation schemes. For example, for two candidate power supply nodes, current allocation ratios of (100%, 0%), (80%, 20%), (60%, 40%), (50%, 50%), (40%, 60%), (20%, 80%), and (0%, 100%) can be generated. The total power loss and voltage drop recovery time of the power supply network under each current allocation scheme are calculated, forming a set of simulation results.

[0034] Based on the simulation results, the optimal current allocation path and phase configuration are selected. A multi-objective optimization function is defined, considering both minimizing the overall power supply network loss and minimizing the voltage transient response time. Pareto front analysis is performed on each scheme in the simulation result set to filter out the non-dominated solution set. In the non-dominated solution set, the optimization tendency emphasizing either power supply efficiency or dynamic response speed is selected based on the current game scenario type. For frame rate-sensitive scenarios such as first-person shooter games, the focus is on dynamic response speed, selecting the scheme with the shortest voltage drop recovery time; for power consumption-sensitive scenarios such as long-duration games, the focus is on power supply efficiency, selecting the scheme with the minimum overall power supply network loss. Combining the optimization tendency with the hardware execution overhead of phase switching, the scheme with the highest comprehensive score is selected as the optimal current allocation strategy.

[0035] A coordinated control instruction set is generated, encompassing voltage regulation module switching frequency adjustment, multi-phase current distribution ratio, and load transient response compensation strategy. Based on the optimal current distribution strategy, the target switching frequency of each voltage regulation module is determined. The switching frequency is reduced under light load to decrease switching losses, and increased under heavy load to improve transient response capability. Based on the multi-phase current distribution ratio, the current output values ​​of each power supply phase are reconfigured, and dynamic balance of load current across multiple phases is achieved by adjusting the pulse width modulation duty cycle of each phase. Based on the load transient response compensation strategy, the reference voltage or feedforward compensation parameters of the voltage regulation module are pre-adjusted. The principle of feedforward compensation is to adjust the control signal of the voltage regulation module in advance when the predicted load current changes, offsetting the voltage drop caused by the predicted load surge. Specifically, based on the predicted current increment ΔI... pred Calculate the feedforward compensation amount ΔV FF = K·ΔI pred , where K is the feedforward gain coefficient, and this compensation is superimposed on the reference voltage of the voltage regulation module.

[0036] The coordinated control instruction set is sent to the power management microcontroller for execution. The power management microcontroller receives the instruction sequence through the system management bus or a dedicated communication interface, parses it, and writes it into the configuration register of the voltage regulation module to complete the real-time adjustment of switching frequency, phase current distribution, and feedforward compensation parameters.

[0037] This invention differs significantly from existing power management methods for gaming PCs in both function and structure, primarily in the following aspects: Load prediction and proactive regulation. Existing technologies employ passive voltage feedback regulation, which suffers from inherent latency. This invention uses a temporal convolutional network and attention mechanism to predict future loads, pre-adjusting the power distribution state before load surges occur, thus achieving a shift from passive response to proactive prediction. Load feature decoupling and differentiated response. Existing technologies have a single dimension of load perception and cannot distinguish between different types of load features, such as graphics rendering, physics calculations, and peripheral interruptions. This invention separates three types of load features through feature decoupling, enabling the power system to configure differentiated response modes according to load type. Power supply network collaborative optimization. Existing technologies control each power supply unit independently, lacking overall coordination. This invention constructs a power distribution topology and simulates current distribution on a graphical model, achieving joint power supply allocation across multiple voltage regulation modules and multiple power supply phases.

[0038] The beneficial technical effects of this invention also include: suppressing voltage transient drops by adjusting the control parameters of the voltage regulation module before load surges occur through load prediction and pre-regulation, significantly reducing the voltage drop amplitude; improving power supply efficiency by reducing the overall power loss of the power supply network through collaborative optimization and dynamic impedance matching of the global power supply network; and enhancing load adaptability by enabling the power system to accurately adapt to different types of gaming loads through load characteristic decoupling and differentiated regulation strategies. In summary, this invention, through the combination of load prediction, characteristic decoupling, and power supply network collaborative optimization technologies, overcomes the limitations of existing power management methods for gaming PCs, providing a more efficient and stable power regulation solution, and greatly improving the power supply performance and system stability of gaming PCs under high load scenarios.

[0039] Example 2 The core inventive concept of this invention lies in achieving proactive power system regulation through load prediction, combined with load characteristic decoupling and power supply network collaborative optimization, to improve the power supply stability and efficiency of gaming PCs. Besides the implementation method described above, which uses temporal convolutional networks and attention mechanisms as its core, this invention can also be implemented through several other different technical paths.

[0040] One specific implementation is based on a Long Short-Term Memory (LSTM) network. This approach uses an LTM network as the core of the load prediction model, which has the advantage of strong modeling ability for long-range dependencies in time series. Specifically, the decoupled load features and game scene context data are input into the LTM network in time series. Its gating mechanism remembers and forgets historical information, outputting the predicted load pressure index for future time windows. The difference between this approach and the previous embodiments is that it uses an LTM network instead of the combination of temporal convolutional networks and attention mechanisms, resulting in a simpler structure and suitability for game scenarios with relatively regular load patterns.

[0041] Another specific implementation is based on reinforcement learning. This approach models the power regulation problem as a reinforcement learning problem, learning the optimal regulation strategy through interaction between the agent and the environment. Specifically, the load prediction and the current power supply network state are used as the state space, the switching frequency adjustment and phase current allocation are used as the action space, and voltage stability and power supply efficiency are used as reward signals. The regulation strategy network is trained using a deep Q-network or a policy gradient method. The difference between this approach and the previous embodiments is that it integrates the prediction and decision-making processes into end-to-end policy learning, avoiding explicit simulation and optimization processes. It is suitable for scenarios with complex and variable load patterns and where it is difficult to establish an accurate physical model.

[0042] Another specific implementation is based on model predictive control. This approach uses a model predictive control framework. Within each control cycle, it predicts the future state based on a load prediction model, obtains the optimal control sequence by solving a constrained optimization problem, and executes only the first control variable. The difference between this approach and the previous embodiments is that it employs a closed-loop rolling optimization control framework, which can continuously correct prediction errors during operation and exhibits strong robustness to model uncertainties and external disturbances.

[0043] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A load prediction-based intelligent regulation method for e-sports power supply, characterized in that, For use in gaming PC systems, the following steps are included: The system collects real-time load data streams and game scene context data during the operation of a gaming PC. The real-time load data streams include CPU core utilization sequences, GPU core utilization sequences, video memory bandwidth utilization waveforms, and / or motherboard power supply phase current fluctuation trajectories. The game scene context data includes game frame rate, game scene tags, and / or peripheral input event frequency. By performing load feature decoupling processing on the real-time load data stream, the load features of graphics rendering, physical computing, and peripheral interrupts are separated. The decoupled load features and the game scene context data are input together into the load prediction model based on temporal convolutional networks and attention mechanisms to generate the predicted load pressure index of each hardware component within the future time window. Based on the motherboard power supply architecture and the physical connection of each hardware component, a power distribution topology is constructed. The nodes of the power distribution topology represent voltage regulation modules and the hardware components connected to them, and the edges represent power transmission paths and their dynamic impedance parameters. Perform current distribution simulation on the power distribution topology, calculate the overall power network loss and voltage transient response time under different current distribution schemes based on the predicted load pressure index, select the optimal current distribution path and phase configuration based on the simulation results, and generate a set of coordinated control instructions including voltage regulation module switching frequency adjustment, multi-phase current distribution ratio and load transient response compensation strategy. The coordinated control instruction set is sent to the power management microcontroller for execution.

2. The method of claim 1, wherein, The load characteristic decoupling process for the real-time load data stream includes: Joint spectrum analysis is performed on the graphics processor core utilization sequence and memory bandwidth utilization waveform to extract graphics rendering cycle features and texture data stream features, and generate the graphics rendering load features. Perform multi-scale wavelet transform on the CPU core utilization sequence to extract rigid body dynamics calculation features and scene loading calculation features in physical calculations, and generate the physical calculation type load features. A time-series correlation analysis is performed on the frequency of peripheral input events and the changes in game frame rate to identify peripheral trigger-type load peaks and generate peripheral interrupt-type load characteristics.

3. The method of claim 1, wherein the method further comprises: The generation of the predicted load stress index for each hardware component within the future time window includes: The graphics rendering load characteristics, physics calculation load characteristics, peripheral interrupt load characteristics, and game scene context data are aligned along the time dimension to form a multi-channel time-series input matrix. A temporal convolutional network is used to extract multi-scale temporal features from the multi-channel temporal input matrix to capture load change patterns over different time spans. By using a multi-head attention mechanism to weightedly fuse the extracted temporal features, key time point features that are strongly correlated with future load peaks are obtained. The weighted and fused features are input into a fully connected prediction network, which outputs the predicted load pressure index of each hardware component within a future predetermined time window.

4. The method of claim 1, wherein the method further comprises: The construction of the power distribution topology includes: Scan the motherboard power management architecture to identify the voltage regulation module, the power supply phase of the central processing unit, the power supply phase of the graphics processing unit, and their respective hardware loads. Measure the dynamic impedance characteristics of each power transmission path under different current loads and establish an impedance-current function model. The voltage regulation module and hardware components are abstracted as nodes, and the node attributes include the predicted load pressure index; The power transmission path is abstracted as a directed edge, and the edge attributes include real-time current value, dynamic impedance parameters and the maximum current carried by the path. Based on the power supply connection relationship between the voltage regulation module and the hardware components, a weighted directed graph structure is established for the power distribution topology.

5. The load prediction-based intelligent regulation method for e-sports power supply according to claim 1, characterized in that, The current distribution simulation performed on the power distribution topology includes: Based on the predicted load pressure index, identify hardware components that will experience a surge in current within a future time window as target nodes for regulation. In the power distribution topology diagram, all voltage regulation module nodes connected to the control target node are identified as candidate power supply nodes. Calculate the power loss and transient response delay along the power transmission path from each candidate power supply node to the control target node; The simulation allocates the current demand of the target node to multiple candidate power supply nodes in different proportions, generating various current allocation schemes. Calculate the total power loss and voltage drop recovery time of the power supply network under each current distribution scheme, and form a set of simulation results.

6. The method of claim 5, wherein the method further comprises: The calculation of power loss and transient response delay along the power transmission path from each candidate power supply node to the control target node includes: Calculate the resistive loss increment on the power transmission path based on the current current value, dynamic impedance parameters, and the current increment to be allocated. Based on the switching frequency of the voltage regulation module and the output capacitor parameters, calculate the transient response time of the output voltage when the current changes. Query historical power regulation records to obtain data on the actual voltage drop amplitude under similar current jump scenarios; The incremental resistive loss, transient response time, and historical voltage drop magnitude are weighted and integrated to generate a comprehensive power supply cost score.

7. The load prediction based intelligent regulation method for e-sports power supply according to claim 1, characterized in that, The selection of the optimal current distribution path and phase configuration based on simulation results includes: A multi-objective optimization function is defined, which simultaneously considers two objectives: minimizing the overall power supply network loss and minimizing the voltage transient response time. Pareto front analysis was performed on each scheme in the simulation results set to screen out the non-dominated solution set; In the non-dominated solution set, the optimization tendency is selected based on the current game scene type, which includes frame rate sensitive scenes and power consumption sensitive scenes. Taking into account the optimization tendency and the hardware execution overhead of phase switching, the scheme with the highest comprehensive score is selected as the optimal current allocation strategy.

8. The method of claim 7, wherein the method further comprises: The generated collaborative control instruction set, which includes voltage regulation module switching frequency adjustment, multi-phase current distribution ratio, and load transient response compensation strategy, includes: Based on the optimal current distribution strategy, determine the target switching frequency of each voltage regulation module; Based on the multi-phase current distribution ratio, the current output value of each power supply phase is reconfigured to achieve dynamic balance of load current among multiple phases. Based on the load transient response compensation strategy, the reference voltage or feedforward compensation parameters of the voltage regulation module are adjusted in advance to offset the voltage drop caused by the predicted load surge. The switching frequency adjustment instructions, current distribution instructions, and transient compensation parameters are encapsulated into a register configuration sequence that can be executed by the power management microcontroller.

9. The load prediction based intelligent regulation method for e-sports power supply according to claim 1, characterized in that, The method further includes: The feedback learning process is optimized, and after executing the coordinated control instruction set, the actual voltage fluctuation curve, power supply efficiency data, and game frame rate stability data are continuously recorded. The actual data is compared with the expected values ​​in the simulation results to calculate the power supply regulation deviation rate; If the power supply regulation deviation rate exceeds the preset threshold, the online update of the load prediction model is triggered, and the weights of the temporal convolutional network and attention mechanism are fine-tuned using the newly collected data. The input characteristics, output commands, and actual effects of each adjustment are stored in the power optimization knowledge base for offline training and optimization of subsequent adjustment strategies.

10. The load prediction based intelligent regulation method for e-sports power supply according to claim 1, characterized in that, The game scene context data also includes: the state of the game engine rendering pipeline, the predicted value of changes in the number of players in multiplayer online games, and the macro operation sequence of e-sports peripherals.

Images