Low-power-consumption dynamic adjustment method and system based on multi-objective optimization decision

Abstract

The invention relates to a low-power-consumption dynamic adjustment method and system based on a multi-objective optimization decision, and the method comprises the steps: generating a perception state signal containing a screen refresh rate demand and a brightness demand through monitoring a user interaction state, a dynamic change rate of display content and ambient light intensity in real time; combining the perceived state signal with central processing unit utilization rate, working frequency and load prediction information collected from system hardware to construct a system state vector; inputting into a pre-trained reinforcement learning model, and outputting a scheduling action including adjusting the working parameters of the central processing unit, the screen refresh rate and the screen brightness by the reinforcement learning model according to a current strategy; executing a scheduling action, estimating the total power consumption of the executed system in real time according to a system power consumption model, detecting system performance and task response delay at the same time, and generating a comprehensive feedback reward signal; and updating a strategy of the reinforcement learning model to maximize long-term accumulated rewards and realize self-adaptive balance of energy efficiency and performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of embedded device power consumption adjustment, more particularly, to a low-power dynamic adjustment method and system based on multi-objective optimization decision. BACKGROUND

[0002] Under the background of continuous improvement of embedded device performance, power management technology is facing severe challenges. As the mainstream power consumption control method, traditional dynamic voltage and frequency scaling (DVFS) has been difficult to meet the fine demand of modern intelligent terminals for energy efficiency due to its extensive management strategy. The technology (DVFS) only relies on historical load data for passive response, and there are significant response delays and misjudgment problems. When the system load changes suddenly, the lagging adjustment of DVFS will lead to a sharp increase in instantaneous power consumption or excessive performance reduction, directly affecting the user experience.

[0003] At the same time, the existing power management framework has obvious subsystem power consumption blind area. The current solution focuses too much on the energy consumption optimization of CPU cores, but ignores the coordinated power consumption management of non-CPU components such as display screens, graphic processors (GPU), high-speed buses and peripheral modules. These subsystems contribute up to 40% of the total system power consumption in multimedia processing and data transmission scenarios, but lack a joint optimization mechanism across components, leading to a bottleneck in overall energy efficiency improvement.

[0004] Moreover, the balance dilemma between performance and power consumption further exacerbates the technical challenge. The traditional method cannot achieve "precise enough" power consumption control, and often appears in the two polarized phenomena of performance overshoot or performance deficiency.

[0005] In addition, the design flaws of the sleep-wake mechanism also restrict the energy efficiency optimization. The existing sleep strategy relies on static prediction models and cannot accurately predict user operation patterns, resulting in frequent sleep-wake cycles. Test data shows that each wake-up process consumes an average of 3-5 millijoules of energy, and the wrong sleep decision increases standby power consumption by 15-20%.

[0006] These technical defects together constitute a common problem in the field of embedded device power management that needs to be broken through, and a low-power dynamic adjustment method and system based on multi-objective optimization decision is needed to solve these problems. SUMMARY

[0007] The technical problem to be solved by the present application is to provide a low-power dynamic adjustment method based on multi-objective optimization decision, and a low-power dynamic adjustment system based on multi-objective optimization decision, in view of the above-mentioned defects of the prior art.

[0008] The technical solution adopted by the present application to solve its technical problems is: A low-power dynamic adjustment method based on multi-objective optimization decision is constructed, wherein the method comprises the steps of: Real-time monitoring of user interaction state, display content dynamic change rate and ambient light intensity, generating a perception state signal containing screen refresh rate demand and brightness demand; The perception state signal is combined with the central processor usage rate, working frequency and load prediction information collected from the system hardware to construct a system state vector; The system state vector is input into a pre-trained reinforcement learning model, and the reinforcement learning model outputs a scheduling action containing adjustment of central processor working parameters, screen refresh rate and screen brightness according to the current strategy; The scheduling action is executed, and the system total power consumption after execution is estimated in real time according to the system power consumption model, while the system performance and task response delay are detected to generate a comprehensive feedback reward signal; According to the system state vector, scheduling action, feedback reward signal and next time system state vector, the strategy of the reinforcement learning model is updated.

[0009] The low-power dynamic adjustment method based on multi-objective optimization decision, wherein in the perception step, the dynamic change rate of the display content The system frame buffer is analyzed in real time.

[0010] The low-power dynamic adjustment method based on multi-objective optimization decision, wherein the specific steps of adjusting the screen refresh rate comprise: Based on the dynamic change rate The target refresh rate is calculated by the following formula : ; Wherein, And are the preset minimum and maximum values of the refresh rate, and k is the adjustment coefficient; When the dynamic change rate Is lower than the set first threshold value, the screen refresh rate is adjusted to 30Hz to 50Hz; when the dynamic change rate Is higher than the set second threshold value, the screen refresh rate is adjusted to 60Hz or above 60Hz.

[0011] The low-power dynamic adjustment method based on multi-objective optimization decision, wherein the user interaction state Is a binary flag, which is marked as valid state when user touch, keyboard, mouse or gaze input is detected, otherwise it is marked as invalid state; The specific steps of adjusting the screen brightness comprise: The average brightness of the display content and user interaction state The target screen brightness L is calculated by the following formula: ; Wherein, and are the minimum and maximum values of the screen brightness allowed by the system respectively, is a preset maximum ambient light reference value, and γ is a brightness response curve parameter.

[0012] The low-power dynamic adjustment method based on multi-objective optimization decision of the present application, wherein when the user interaction state is an invalid state, the target screen brightness L is set to the minimum brightness or a preset energy-saving brightness value.

[0013] The low-power dynamic adjustment method based on multi-objective optimization decision of the present application, wherein when the average brightness of the display content is higher than a preset brightness threshold, the high dynamic range display technology is enabled to maintain the visual perception of the user to the picture brightness while the backlight brightness is lowered.

[0014] The low-power dynamic adjustment method based on multi-objective optimization decision of the present application, wherein the total system power consumption of the system power consumption model is estimated according to the following formula: ; Wherein, , , , and are the power consumptions of the central processor, the graphic processor, the memory, the display module and other subsystems respectively.

[0015] The low-power dynamic adjustment method based on multi-objective optimization decision of the present application, wherein the model of the power consumption of the central processor is: ; Wherein, is the switched capacitor, is the working voltage, is the working frequency, is the static power consumption; the model is calibrated in real time by reading the power management unit register built in the chip; The model of the power consumption of the graphic processor is: ; Wherein, is the GPU shader core power consumption coefficient, which is proportional to the square of the core voltage; is the GPU shader core working frequency; is the GPU shader core utilization, which reflects the computing load; is the GPU memory interface power consumption coefficient; is the GPU memory bandwidth utilization; is the static power consumption of the GPU; is the power consumption of the memory The model is: ; wherein, and are the power consumption coefficients of DRAM read and write operations respectively; is the working frequency of the DRAM; and represent the read bandwidth utilization and the write bandwidth utilization respectively; is the background power consumption of the memory; The power consumption of the display module The model is: ; ; ; wherein, is the driving power consumption of the display panel (such as LCD); is the energy efficiency coefficient of the panel; is the actual refresh rate of the screen; is the proportion of the active refresh area to the total area; is the average pixel brightness value of the display content; is the basic power consumption constant of the panel bias and driving circuit; is the backlight module power consumption; is the energy efficiency coefficient of the backlight module; is the backlight brightness level; is the total area of the panel; The power consumption of other subsystems The model is: ; wherein, is the power consumption of the on-chip interconnection bus; is the power consumption of the memory controller; is the power consumption of various peripheral modules.

[0016] The low-power dynamic adjustment method based on multi-target optimization decision in the application, wherein the feedback reward signal , for evaluating the comprehensive effect of performing action under state , which is formulated as follows: ; wherein, , , are weight coefficients, is the total power consumption of the system, is a performance indicator function, taking the value of 1 when the performance indicator is lower than a preset threshold Thresh, otherwise 0, is the task response delay; State is represented by the formula: ; wherein, is the CPU usage rate, is the current CPU working frequency, is the predicted future load, is the user activity state, is the ambient light intensity; Action is the control decision made by the scheduler at time t according to the current state, for dynamically adjusting system resources, which is formally represented as: ; wherein, is the adjustment of the CPU working frequency, is the dynamic switching of the CPU core, is the adjustment of the screen brightness.

[0017] A low-power dynamic adjustment system based on multi-objective optimization decision, for implementing the low-power dynamic adjustment method based on multi-objective optimization decision as described above, wherein the system comprises a state sensing unit, a system state construction unit, a scheduling decision unit, an action execution and feedback unit, and a model optimization unit: The state sensing unit is configured to monitor the user interaction state, the dynamic change rate of the display content, and the ambient light intensity in real time, and generate a sensing state signal containing screen refresh rate demand and brightness demand; The system state construction unit is configured to combine the sensing state signal with the central processor usage rate, working frequency, and load prediction information collected from the system hardware, to construct a system state vector; The scheduling decision unit is configured to input the system state vector into a pre-trained reinforcement learning model, and output a scheduling action containing adjustment of the central processor working parameters, screen refresh rate, and screen brightness from the reinforcement learning model according to the current strategy; The action execution and feedback unit is used to execute scheduling actions, estimate the total system power consumption after execution in real time according to the system power consumption model, detect system performance and task response latency, and generate a comprehensive feedback reward signal. The model optimization unit is used to update the strategy of the reinforcement learning model based on the system state vector, scheduling actions, feedback reward signals and the system state vector at the next moment, so as to maximize the long-term cumulative reward and achieve an adaptive balance between energy efficiency and performance.

[0018] The beneficial effects of this invention are as follows: By applying the method of this application, a perceived state signal containing screen refresh rate and brightness requirements is generated through real-time monitoring of user interaction status, dynamic change rate of displayed content, and ambient light intensity; the perceived state signal is combined with CPU utilization, operating frequency, and load prediction information collected from system hardware to construct a system state vector; the system state vector is input into a pre-trained reinforcement learning model, which outputs scheduling actions including adjusting CPU operating parameters, screen refresh rate, and screen brightness according to the current strategy; the scheduling actions are executed, and the total system power consumption after execution is estimated in real time according to the system power consumption model, while system performance and task response latency are detected to generate a comprehensive feedback reward signal; the strategy of the reinforcement learning model is updated based on the system state vector, scheduling actions, feedback reward signal, and the system state vector at the next moment to maximize long-term cumulative rewards and achieve an adaptive balance between energy efficiency and performance. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort: Figure 1 This is a flowchart of a low-power dynamic adjustment method based on multi-objective optimization decision-making according to a preferred embodiment of the present invention. Figure 2 This is a block diagram of a low-power dynamic adjustment system based on multi-objective optimization decision-making, which is a preferred embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, a clear and complete description will be provided below in conjunction with the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0021] The preferred embodiment of the present invention is a low-power dynamic adjustment method and system based on multi-objective optimization decision-making, such as... Figure 1 As shown, the steps include: S01: Real-time monitoring of user interaction status, dynamic change rate of displayed content, and ambient light intensity to generate a perception status signal that includes screen refresh rate requirements and brightness requirements; Dynamic change rate of displayed content This is obtained through real-time analysis of the system frame buffer; The specific steps for adjusting the screen refresh rate include: Based on dynamic rate of change The target refresh rate is calculated using the following formula. : ; in, and These are the preset minimum and maximum refresh rates, where k is the adjustment coefficient. When the dynamic rate of change When the refresh rate is below the set first threshold, adjust the screen refresh rate to 30Hz to 50Hz; when the dynamic change rate... When the refresh rate exceeds the set second threshold, adjust the screen refresh rate to 60Hz or higher. User interaction status It is a binary flag; it is marked as valid when user touch, keyboard and mouse or gaze input is detected, and otherwise it is marked as invalid. The specific steps for adjusting screen brightness include: Overall ambient illuminance E, average brightness of displayed content and user interaction status The target screen brightness L is calculated using the following formula: ; in, and These are the minimum and maximum screen brightness allowed by the system, respectively. γ is the preset maximum ambient light reference value, and γ is the parameter of the brightness response curve. When user interaction status When in an invalid state, the target screen brightness L is set to the minimum brightness value. Or a preset energy-saving brightness value; When the average brightness of the displayed content When the brightness exceeds the preset threshold, high dynamic range display technology is activated to maintain the user's visual perception of the screen brightness while reducing the backlight brightness.

[0022] S02: Combine the sensed state signals with the CPU utilization, operating frequency and load prediction information collected from the system hardware to construct the system state vector; S03: Input the system state vector into the pre-trained reinforcement learning model, and the reinforcement learning model outputs scheduling actions including adjusting the central processing unit working parameters, screen refresh rate and screen brightness according to the current policy; S04: Execute scheduling actions and estimate the total system power consumption after execution in real time based on the system power consumption model. At the same time, detect system performance and task response latency, and generate a comprehensive feedback reward signal. The total system power consumption model Estimate using the following formula: ; in, , , , and These are the power consumption components of the central processing unit, graphics processing unit, memory, display module, and other subsystems; the scheduler performs joint optimization by monitoring and predicting changes in these components. Central Processing Unit Power Consumption The model is: ; in, For switched capacitors, which are related to chip manufacturing process and physical design, they are constants. The operating voltage of the CPU voltage domain, measured in volts (V); related to frequency. Typically, the relationship is linear (adjusted via DVFS). This represents the average operating frequency of the CPU cluster, measured in Hertz (Hz). The static power consumption of the CPU cluster is mainly caused by leakage current, which is related to temperature and voltage and is relatively fixed at a given operating point. Power consumption of graphics processor The model is: ; in, This is the GPU shader core power consumption coefficient, which is proportional to the square of the core voltage. This refers to the operating frequency of the GPU Shader core; GPU shader core utilization (between 0 and 1), reflecting the computational load; This refers to the power consumption factor of the GPU memory interface. GPU memory bandwidth utilization (unit: GB / s); This refers to the static power consumption of the GPU. Memory power consumption The model is: ; in, and These are the power consumption coefficients for DRAM read and write operations (read and write operations consume different amounts of energy). This refers to the operating frequency of DRAM (such as the data rate of LPDDR4). and These represent read bandwidth utilization and write bandwidth utilization (between 0 and 1), respectively. This refers to the background power consumption of memory, including refresh power consumption and standby power consumption; Power consumption of the display module The model is: ; ; ; in, Power consumption for driving display panels (such as LCDs); The panel's energy efficiency coefficient; The actual screen refresh rate (Hz) can be dynamically adjusted to significantly save energy. The proportion of the active refresh area to the total area; partial refresh technology saves energy by reducing this value. The average pixel brightness value of the displayed content (normalized from 0 to 1); This is the fundamental power consumption constant for panel bias and drive circuitry; Power consumption of the backlight module; The energy efficiency coefficient of the backlight module; For backlight brightness levels (0-1 normalized), power consumption is usually proportional to the square of the brightness. This represents the total area of ​​the panel. Power consumption of other subsystems The model is: ; in, Power consumption for on-chip interconnect buses (such as NOC) is related to bus frequency and load; The power consumption of the storage controller (such as eMMC / SD) is related to read and write activity; For the power consumption of various peripheral modules (such as audio codecs, video input interfaces, network controllers, etc.), each module has independent enable / load control; These formulas together constitute a real-time, predictable digital twin model of system power consumption. In actual operation: 1. Parameter calibration: The coefficients (C, α, β, γ, δ, ε, η, etc.) are obtained by joint calibration of the hardware performance counter (PMU) and precision power meter at the factory.

[0023] 2. Real-time input: The scheduler obtains dynamic variables such as f, V, U, B, R, L, etc. from the kernel and drivers in real time.

[0024] 3. Prediction and Decision Making: By combining this model with a load predictor, the scheduler (such as the Model Predictive Controller, MPC) can pre-evaluate the power consumption and performance under different scheduling strategies, thereby selecting the globally optimal energy efficiency point that meets performance constraints.

[0025] S05: Update the strategy of the reinforcement learning model based on the system state vector, scheduling action, feedback reward signal and the system state vector at the next moment; Feedback reward signal Used to evaluate the state Next action The overall effect is expressed by the following formula: ; in, , , All are weighting coefficients. The total power consumption of the system is This is a performance indicator function; it takes a value of 1 when the performance metric is below a preset threshold (Thresh), and 0 otherwise. Due to task response delay; state The formula is used to indicate: ; in, For CPU utilization, The current operating frequency of the CPU. For the predicted future load, For user activity status, Ambient light intensity; action The control decision made by the scheduler at time t based on the current state, used to dynamically adjust system resources, is formally represented as: ; in, To adjust the CPU operating frequency, To dynamically switch the CPU core, To adjust screen brightness.

[0026] The application of the method of this application has the following effects: Overall energy efficiency ratio optimization: Under the same workload, the system's total energy efficiency (performance / watt) is improved by more than 30%.

[0027] Significantly extended battery life: Under moderate usage scenarios with continuous interaction, the overall battery life is extended by 20-35%.

[0028] Performance Guarantee: Ensure that the UI rendering frame rate is no less than 60fps, and that key user experience metrics such as application startup latency are not affected.

[0029] Smooth experience: Avoid momentary stuttering caused by frequency reduction or core shutdown.

[0030] A low-power dynamic adjustment system based on multi-objective optimization decision-making is provided to implement the low-power dynamic adjustment method based on multi-objective optimization decision-making as described above. The system includes a state sensing unit 300, a system state construction unit 301, a scheduling decision-making unit 302, an action execution and feedback unit 303, and a model optimization unit 304. The state perception unit 300 is used to monitor the user interaction state, the dynamic change rate of the displayed content, and the ambient light intensity in real time, and generate a perception state signal that includes screen refresh rate requirements and brightness requirements. The system state construction unit 301 is used to combine the sensed state signal with the CPU utilization, operating frequency and load prediction information collected from the system hardware to construct the system state vector. The scheduling decision unit 302 is used to input the system state vector into the pre-trained reinforcement learning model, and the reinforcement learning model outputs scheduling actions including adjusting the central processing unit working parameters, screen refresh rate and screen brightness according to the current policy. The action execution and feedback unit 303 is used to execute scheduling actions, estimate the total system power consumption after execution in real time according to the system power consumption model, and at the same time detect system performance and task response latency to generate a comprehensive feedback reward signal. The model optimization unit 304 is used to update the strategy of the reinforcement learning model based on the system state vector, scheduling action, feedback reward signal and the system state vector at the next moment; The system using this application has the following effects: Overall energy efficiency ratio optimization: Under the same workload, the system's total energy efficiency (performance / watt) is improved by more than 30%.

[0031] Significantly extended battery life: Under moderate usage scenarios with continuous interaction, the overall battery life is extended by 20-35%.

[0032] Performance Guarantee: Ensure that the UI rendering frame rate is no less than 60fps, and that key user experience metrics such as application startup latency are not affected.

[0033] Smooth experience: Avoid momentary stuttering caused by frequency reduction or core shutdown.

[0034] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A low-power dynamic adjustment method based on multi-objective optimization decision-making, characterized in that, The method includes the following steps: Real-time monitoring of user interaction status, dynamic change rate of displayed content, and ambient light intensity generates a perception status signal that includes screen refresh rate requirements and brightness requirements. The system state vector is constructed by combining the sensed state signals with information on CPU utilization, operating frequency, and load prediction collected from the system hardware. The system state vector is input into a pre-trained reinforcement learning model, which then outputs scheduling actions based on the current policy, including adjusting the CPU operating parameters, screen refresh rate, and screen brightness. Execute scheduling actions and estimate the total system power consumption after execution in real time based on the system power consumption model. At the same time, detect system performance and task response latency, and generate a comprehensive feedback reward signal. The policy of the reinforcement learning model is updated based on the system state vector, scheduling actions, feedback reward signals, and the system state vector at the next moment.

2. The low-power dynamic adjustment method based on multi-objective optimization decision-making according to claim 1, characterized in that, The dynamic change rate of the displayed content This is obtained through real-time analysis of the system frame buffer.

3. The low-power dynamic adjustment method based on multi-objective optimization decision-making according to claim 2, characterized in that, The specific steps for adjusting the screen refresh rate include: Based on the dynamic rate of change The target refresh rate is calculated using the following formula. : ； in, and These are the preset minimum and maximum refresh rates, where k is the adjustment coefficient. When the dynamic rate of change When the refresh rate is below the set first threshold, adjust the screen refresh rate to 30Hz to 50Hz; when the dynamic change rate... If the refresh rate exceeds the set second threshold, adjust the screen refresh rate to 60Hz or higher.

4. The low-power dynamic adjustment method based on multi-objective optimization decision-making according to claim 1, characterized in that, The user interaction state It is a binary flag; it is marked as valid when user touch, keyboard and mouse or gaze input is detected, and otherwise it is marked as invalid. The specific steps for adjusting screen brightness include: Overall ambient illuminance E, average brightness of displayed content and user interaction status The target screen brightness L is calculated using the following formula: ； in, and These are the minimum and maximum screen brightness allowed by the system, respectively. γ is the preset maximum ambient light reference value, and γ is the parameter of the brightness response curve.

5. The low-power dynamic adjustment method based on multi-objective optimization decision-making according to claim 4, characterized in that, When user interaction status When in an invalid state, the target screen brightness L is set to the minimum brightness value. Or a preset energy-saving brightness value.

6. The low-power dynamic adjustment method based on multi-objective optimization decision-making according to claim 4, characterized in that, When the average brightness of the displayed content When the brightness exceeds the preset threshold, high dynamic range display technology is activated to maintain the user's visual perception of the screen brightness while reducing the backlight brightness.

7. The low-power dynamic adjustment method based on multi-objective optimization decision-making according to claim 1, characterized in that, The system power consumption model represents the total system power consumption. Estimate using the following formula: ； in, , , , and These are the power consumption of the central processing unit, graphics processing unit, memory, display module, and other subsystems, respectively.

8. The low-power dynamic adjustment method based on multi-objective optimization decision-making according to claim 7, characterized in that, The power consumption of the central processing unit The model is: ； in, For switched capacitors, Operating voltage For operating frequency, This represents static power consumption; the model calibrates parameters in real time by reading the power management unit register built into the chip. Power consumption of graphics processor The model is: ； in, This is the GPU shader core power consumption coefficient, which is proportional to the square of the core voltage. This refers to the operating frequency of the GPU Shader core. GPU shader core utilization reflects the computational load; This refers to the power consumption factor of the GPU memory interface. This refers to the GPU's memory bandwidth utilization. This refers to the static power consumption of the GPU. Memory power consumption The model is: ； in, and These are the power consumption coefficients for DRAM read and write operations, respectively. Refers to the operating frequency of DRAM; and These represent read bandwidth utilization and write bandwidth utilization, respectively. This refers to the background power consumption of the memory; Power consumption of the display module The model is: ； ； ； in, Power consumption for display panel drivers; The panel's energy efficiency coefficient; This refers to the actual screen refresh rate. The percentage of the total area to be refreshed for the activity; The average pixel brightness value of the displayed content; This is the fundamental power consumption constant for panel bias and drive circuitry; Power consumption of the backlight module; The energy efficiency coefficient of the backlight module; Backlight brightness level; This represents the total area of ​​the panel. Power consumption of other subsystems The model is: ； in, The power consumption of the on-chip interconnect bus; The power consumption of the storage controller; Power consumption for various peripheral modules.

9. The low-power dynamic adjustment method based on multi-objective optimization decision-making according to claim 7, characterized in that, The feedback reward signal Used to evaluate the state Next action The overall effect is expressed by the following formula: ； in, , , All are weighting coefficients. The total power consumption of the system is This is a performance indicator function; it takes a value of 1 when the performance metric is below a preset threshold (Thresh), and 0 otherwise. Due to task response delay; state The formula is used to indicate: ； in, CPU utilization This is the current operating frequency of the CPU. For the predicted future load, For user activity status, Ambient light intensity; action The control decision made by the scheduler at time t based on the current state, used to dynamically adjust system resources, is formally represented as: ； in, To adjust the CPU operating frequency, To dynamically switch the CPU core, To adjust screen brightness.

10. A low-power dynamic adjustment system based on multi-objective optimization decision-making, used to implement the low-power dynamic adjustment method based on multi-objective optimization decision-making as described in any one of claims 1-9, characterized in that, The system includes a state awareness unit, a system state construction unit, a scheduling decision unit, an action execution and feedback unit, and a model optimization unit. The state perception unit is used to monitor the user interaction state, the dynamic change rate of the displayed content, and the ambient light intensity in real time, and generate a perception state signal that includes screen refresh rate requirements and brightness requirements. The system state construction unit is used to combine the sensed state signals with the CPU utilization, operating frequency and load prediction information collected from the system hardware to construct a system state vector. The scheduling decision unit is used to input the system state vector into the pre-trained reinforcement learning model, and the reinforcement learning model outputs scheduling actions including adjusting the central processing unit working parameters, screen refresh rate and screen brightness according to the current policy. The action execution and feedback unit is used to execute scheduling actions, estimate the total system power consumption after execution in real time according to the system power consumption model, detect system performance and task response latency, and generate a comprehensive feedback reward signal. The model optimization unit is used to update the strategy of the reinforcement learning model based on the system state vector, scheduling action, feedback reward signal and the system state vector at the next moment.

Images