Equipment battery consumption optimization management method based on intelligent control

Abstract

The invention discloses an equipment battery consumption optimization management method based on intelligent control, and relates to the technical field of battery consumption optimization, and the method comprises the steps: taking a modular power domain as a substrate, and building a reinforcement learning model based on multiple agents; time step lengths are divided, state data of the modular power source domain are obtained under each time step, and action decision making is carried out; designing a reward function; obtaining a corresponding action decision value through the state data, the action decision and the reward function of the modular power domain, and constructing a global value network and a global strategy network after repeatedly iterating for a plurality of time steps; and updating the intelligent agent in real time, and dynamically adjusting the states of different modular power supply domains through a global strategy network. According to the method, different modular power domains are modeled into a plurality of agents, a multi-agent reinforcement learning framework is introduced, the problems of lack of coordination among modules and optimization target dispersion in a traditional method are solved, and the overall energy efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of battery consumption optimization, and specifically to a method for optimizing the management of device battery consumption based on intelligent control. Background Art

[0002] With the widespread application of computer devices, as the core energy system, the management and optimization of batteries have become increasingly prominent. Traditional battery management systems (BMS) usually adopt fixed rules for control, lacking dynamic adaptability and global coordination capabilities, and it is difficult to balance power supply performance, battery safety, and service life. Especially in the context of the increasing popularity of modular power supply structures, the state differences among multiple power modules are significant, and how to achieve efficient energy consumption scheduling and coordinated control has become a technical problem. In recent years, the optimization ability of reinforcement learning, especially multi-agent reinforcement learning, in complex systems has received wide attention, which can provide an adaptive, distributed, and multi-objective optimization solution for battery management. Secondly, based on MARL, it can effectively handle distributed resource optimization problems;

[0003] Traditional optimization of device battery power supply mostly adopts a passive response mode, that is, energy-saving measures are only started when the device detects low battery power or high load. This method has obvious hysteresis. Therefore, there is an urgent need for an active and forward-looking optimization method for battery power supply to optimize the battery life without affecting the user experience. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for optimizing the management of device battery consumption based on intelligent control to solve the problems raised in the prior art.

[0005] To achieve the above purpose, the present invention provides the following technical solution: A method for optimizing the management of device battery consumption based on intelligent control, and the device battery consumption optimization management method specifically includes the following steps:

[0006] Step S100: Based on multi-agent, establish a reinforcement learning model with the modular power domain as the basis; including: regarding different modular power domains in the battery power supply as agents, regarding the battery power supply as a global policy network, and the agents having independent local policy networks and local value networks;

[0007] Step S200: Divide the time step. At each time step, the agents respectively obtain the state data of the modular power domain and make action decisions through the independent local policy network;

[0008] Step S300: Design a reward function based on the change amount of the state data of the modular power domain after executing the action decision;

[0009] Step S400: Obtain the corresponding action decision value through the state data, action decision, and reward function of the modular power domain. After the agent iteratively repeats for several time steps, construct a global value network with the action decision value, and construct a global policy network with the state data and action decisions of all modular power domains;

[0010] Step S500: Update the agent in real time, and dynamically adjust the states of different modular power domains through the global policy network.

[0011] In step S100, based on the modular power domain as the basis, establish a reinforcement learning model based on multi-agent; including: taking different modular power domains in the battery power supply as agents, taking the battery power supply as the global policy network, and the agent has an independent local policy network and local value network. Specifically:

[0012] Take different modular power domains in the battery power supply as agent A i , i = 1, 2, …, N, where N represents the total number of modular power domains; i represents the number identifier of the modular power domain;

[0013] Each agent includes:

[0014] Local policy network π θi (a i ∣s i ); The input is the state s i , and the output is the action a i ; θi represents the first network parameter;

[0015] Local value network Q Φi (s i , a i ); The local value network Q Φi (s i , a i ) is used to evaluate the value of the state-action pair; Φi represents the second network parameter;

[0016] Take the battery power supply as the global policy network Π Θ (g|s); where g represents the global action vector; s represents the concatenation of all agent states; Θ represents the global network parameter;

[0017] The global policy network takes the global state vector as the input and outputs a coordination instruction.

[0018] Furthermore,

[0019] s = [s1, s2, …, sN], which is the concatenation of all agent states, (dimension N×d s );

[0020] g is the global action vector (dimension N×d a ), which is used to coordinate each module;

[0021] Furthermore,

[0022] Define the state space: s i =[V i , I i , T i , SOC i , SOH i ; where, V i represents the voltage value; I i represents the current value; T i represents the temperature value; SOC i represents the state of charge; SOH i represents the state of health;

[0023] Among them, the dimension d s of the state space is 5 - dimensional;

[0024] V i represents the voltage value: (sensor measurement value, unit: V);

[0025] I i represents the current value: (sensor measurement value, unit: A);

[0026] T i represents the temperature value: (sensor measurement value, unit: °C);

[0027] SOC i represents the state of charge: (calculated by Coulomb counting method, range [0, 1]);

[0028] SOH i represents the state of health: state of health (calculated by degradation model, range [0, 1]);

[0029] Define the action space: a i =[ΔI i , B i ;

[0030] Among them, ΔI i represents the adjustment amount of the supply current; (normalized value [-1, 1]); B i represents the heat dissipation control coefficient. (range [0, 1]);

[0031] In step S200, divide the time step. At each time step, the agent respectively obtains the state data of the modular power domain and makes action decisions through the independent local policy network. Specifically:

[0032] Step S201: Set a fixed time step Δt, and the discrete time step sequence is t = k * Δt; where k represents the number of iterations and k is a positive integer; make a decision at each time step t.

[0033] Step S202: At time step t, each agent synchronously collects the state s i ;

[0034] Step S203: The local policy network calculates the action, and the calculation formula is expressed as: a i (t)=π θi (a i |s i ) + ε t ; where a i (t) represents the action at time step t; ε t represents the exploration noise term;

[0035] Furthermore, ε t ~N(0, σ 2 );

[0036] The noise term is fitted by a normal distribution and introduced during training, and the initial value of σ is 0;

[0037] Step S204: Execute the action, including:

[0038] Calculate the current value at the next time step based on the current current value and the supply current adjustment amount;

[0039] Calculate the actual heat dissipation power according to the heat dissipation control coefficient.

[0040] Furthermore,

[0041] I i (t + 1)=I i (t)+K_i * ΔI i ; K_i represents the current conversion coefficient; I i (t) represents the current value at time step t; I i (t + 1) represents the current value at time step t + 1; P_cool i =B i *P_max; where P_cool i represents the actual heat dissipation power; P_max represents the maximum heat dissipation power;

[0042] In step S300, based on the change amount of the state data of the modular power supply domain after executing the action decision, design a reward function, specifically:

[0043] The reward function is based on the change amount of the state data Δs i =s i (t + 1)-si Design at time step (t);

[0044] Aims to balance performance, safety, and lifespan;

[0045] Agent A i Reward r at time step t i Is characterized as:

[0046] r i (t) = w1 * R balance + w2 * R safety + w3 * R efficiency ; where R balance Represents the SOC balance reward term; R safety Represents the temperature safety reward term; R efficiency Represents the efficiency optimization reward term; w1, w2, and w3 are the corresponding reward weight coefficients, satisfying w1 + w2 + w3 = 1.

[0047] In step S400, the corresponding action decision value is obtained through the state data, action decision, and reward function of the modular power domain. After the agent iterates through several time steps, the action decision value forms a global value network, and the state data and action decisions of all modular power domains form a global policy network. Specifically:

[0048] Step S401: Obtain the corresponding action decision value through the state data, action decision, and reward function of the modular power domain;

[0049] Step S402: Iterate through several time steps;

[0050] Step S403: Form a global value network with the action decision value. The global value network is q(s i , a i , r i , Φi); where Φi represents the second network parameter of the local value network in the i-th modular power domain;

[0051] Step S404: Form a global policy network with the state data and action decisions of all modular power domains.

[0052] Furthermore,

[0053] The parameters of the local policy network are represented as: s i ∈R ds -> a i ∈[-1, 1] da ;

[0054] The parameters of the global policy network are represented as: s ∈ R N*ds -> g ∈ R N*da; R represents a dimension identifier;

[0055] In step S500, the agent is updated in real time, and the states of different modular power domains are dynamically adjusted through a global policy network. Specifically:

[0056] Collect the first network parameters θ1, θ2,..., θN in the current local policy network of all agents, and calculate the global policy network parameter Θ through federated averaging, Θ = (θ1 + θ2 + … + θN) / N; realize the real-time update of the local policy network of the agent;

[0057] By calculating the value loss, obtain the value loss gradient, and update the global value network according to the value loss gradient;

[0058] Dynamically adjust the supply current and heat dissipation power of each modular power domain through the real-time updated global policy network.

[0059] Furthermore,

[0060] Set the SOC balance reward term according to the magnitude of the deviation of the SOC change of the modular power domain from the global average;

[0061] Adjust the cooling power according to a preset heat dissipation control coefficient, and design a temperature safety reward term.

[0062] Furthermore,

[0063] The dynamic adjustment of the supply current and heat dissipation power of each modular power domain through the real-time updated global policy network includes:

[0064] The supply current and heat dissipation power are independently controlled respectively.

[0065] Compared with the prior art, the beneficial effects of the present invention are:

[0066] 1. By modeling different modular power domains as multiple agents and introducing a multi-agent reinforcement learning framework, the present invention solves the problems of lack of coordination between modules and scattered optimization objectives in traditional methods, and improves the overall energy efficiency;

[0067] 2. By designing a reward function and comprehensively considering key factors such as current regulation, heat dissipation control, SOC balance, temperature safety and efficiency improvement, the present invention realizes the dynamic balance among power supply performance, operation safety and battery life;

[0068] 3. By introducing a local policy network and a global policy network and combining a reinforcement learning mechanism, the system can continuously learn and optimize decision-making strategies during operation, and adapt to complex working conditions such as load changes and battery aging;

[0069] 4. By intelligently regulating the power supply current and heat dissipation power, the present invention realizes refined energy management, not only improving energy efficiency, but also effectively reducing the risk of failures caused by over-discharge and overheating, thereby reducing the maintenance frequency and extending the service life of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a schematic diagram of the steps of a method for optimizing the battery consumption management of a device based on intelligent control according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0071] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0072] Embodiment: As Figure 1 shown, the present invention provides a technical solution, a method for optimizing the battery consumption management of a device based on intelligent control, and the method for optimizing the battery consumption management of the device specifically includes the following steps:

[0073] Step S100: Based on the modular power domain as the basis, establish a reinforcement learning model based on multiple agents; including: regarding different modular power domains in the battery power supply as agents, regarding the battery power supply as a global policy network, and the agents have independent local policy networks and local value networks;

[0074] Step S200: Divide the time step length. At each time step, the agents respectively obtain the state data of the modular power domain and make action decisions through the independent local policy network;

[0075] Step S300: Design a reward function based on the change amount of the state data of the modular power domain after executing the action decision;

[0076] [[ID=NO=28]]Step S400: Obtain the corresponding action decision value through the state data, action decision and reward function of the modular power domain. After the agents repeat and iterate for several time steps, form a global value network with the action decision value, and form a global policy network with the state data and action decisions of all modular power domains;

[0077] Step S500: Update the agents in real time and dynamically adjust the states of different modular power domains through the global policy network.

[0078] In step S100, a reinforcement learning model is established based on the modular power domain and multiple agents. The model includes: using different modular power domains in the battery power source as agents, using the battery power source as the global policy network, and having independent local policy networks and local value networks for the agents. Specifically,

[0079] Different modular power domains in the battery power supply are used as intelligent agents A i , i=1, 2, …, N, N represents the total number of modular power domains; i represents the number identifier of the modular power domain;

[0080] Each agent consists of:

[0081] Local policy network π θi (a i ∣s i ); input is state s i , the output is action a i ;θi represents the first network parameter;

[0082] Local value network Q Φi (s i ,a i ); the local value network Q Φi (s i ,a i ) is used to evaluate the value of the state-action pair; Φi represents the second network parameter;

[0083] Using battery power as a global policy network π Θ (g|s); where g represents the global action vector; s represents the concatenation of all agent states; Θ represents the global network parameters;

[0084] The global policy network inputs the global state vector and outputs coordination instructions.

[0085] Further,

[0086] s=[s1,s2,…,sN], is the concatenation of all agent states (dimension N×d s );

[0087] g is the global action vector (dimension N×d a ), used to coordinate various modules;

[0088] Further,

[0089] Define the state space: s i =[V i ,I i ,T i ,SOC i ,SOH i ]; where Vi Represents the voltage value; I i Represents the current value; T i Represents the temperature value; SOC i Represents the state of charge; SOH i Represents the state of health;

[0090] Among them, the dimension d of the state space s is 5 - dimensional;

[0091] V i Represents the voltage value: (sensor measurement value, unit: V);

[0092] I i Represents the current value: (sensor measurement value, unit: A);

[0093] T i Represents the temperature value: (sensor measurement value, unit: °C);

[0094] SOC i Represents the state of charge: (calculated by Coulomb counting method, range [0,1]);

[0095] SOH i Represents the state of health: state of health (calculated by degradation model, range [0,1]);

[0096] Define the action space: a i =[ΔI i , B i ;

[0097] Among them, ΔI i Represents the adjustment amount of the supply current; (normalized value [-1,1]); B i Represents the heat dissipation control coefficient. (range [0,1]);

[0098] In step S200, divide the time step. At each time step, the agent respectively obtains the state data of the modular power domain and makes an action decision through an independent local policy network. Specifically:

[0099] [[ID=6l]]Step S201: Set a fixed time step Δt, and the discrete time step sequence is t = k * Δt; where k represents the iteration number and k is a positive integer; make a decision at each time step t;

[0100] Step S202: At time step t, each agent synchronously collects the state s i ;

[0101] Step S203: The local policy network calculates the action, and the calculation formula is represented as: a i (t)=π θi (a i|s i ) + ε t ; where a i (t) represents the action at time step t; ε t represents the exploration noise term;

[0102] Furthermore, ε t ~ N(0, σ 2 );

[0103] The noise term is fitted by a normal distribution and introduced during training, with the initial value of σ being 0;

[0104] Step S204, action execution, includes:

[0105] Calculate the current value at the next time step based on the current current value and the supply current adjustment;

[0106] Calculate the actual heat dissipation power according to the heat dissipation control coefficient.

[0107] Furthermore,

[0108] I i (t + 1) = I i (t) + K_i * ΔI i ; K_i represents the current conversion coefficient; I i (t) represents the current value at time step t; I i (t + 1) represents the current value at time step t + 1; P_cool i = B i * P_max; where P_cool i represents the actual heat dissipation power; P_max represents the maximum heat dissipation power;

[0109] The current conversion coefficient K_i is used to map the normalized value [-1, 1] of the supply current adjustment to the actual current range [I min , I max ;

[0110] The heat dissipation control coefficient is used to control the power of the cooling system;

[0111] In step S300, based on the change amount of the state data of the modular power domain after the action decision is executed, a reward function is designed, specifically:

[0112] The reward function is designed based on the state data change amount Δs i = s i (t + 1) - s i (t);

[0113] The goal is to balance performance, safety, and lifespan;

[0114] Agent Ai The reward r at time step t i (t) is characterized as:

[0115] r i (t) = w1 * R balance + w2 * R safety + w3 * R efficiency ; where R balance represents the SOC balance reward term; R safety represents the temperature safety reward term; R efficiency represents the efficiency optimization reward term; w1, w2, and w3 are the corresponding reward weight coefficients respectively, satisfying w1 + w2 + w3 = 1.

[0116] Among them, if the SOC change of a certain modular power supply domain deviates from the global average value (the larger the absolute value), the reward value is lower (more negative), forcing the agent to adjust the power supply current of this modular power supply domain to narrow the difference from the average value. To avoid over-powering some modular power supply domains (reduce resource waste and extend service life) and achieve balanced utilization; a calculation method for designing the SOC balance reward term is:

[0117] R balance = -|ΔSOC i (t) - Σ N ΔSOC i (t) / N|; where ΔSOC i (t) represents the SOC change amount of the i-th modular power supply domain at time step t; Σ N ΔSOC i (t) / N represents the average value of the SOC change amounts of all N modular power supply domains, reflecting the overall power supply trend of the device;

[0118] To prevent safety risks caused by excessive temperature rise in the modular power supply domain (such as overheating damage or fire), adjust the cooling power through the heat dissipation control coefficient. When Δt i (t) > α, the faster the temperature rises, the heavier the penalty. A calculation method for obtaining the temperature safety reward term is:

[0119] R safety = -max(0, Δt i (t) - α); where Δt i (t) represents the temperature change amount of the i-th modular power supply domain at time step t; α represents the temperature rise safety threshold, that is, the maximum allowable single-step temperature rise;

[0120] If the voltage rises and the current is small, the product is positive and the reward value is higher, encouraging an efficient operating point of "high voltage, small current" (such as low-loss operation close to full charge). If the voltage drops or the current is too large, the product is negative, punishing the low-efficiency state (such as increased internal resistance loss caused by large current discharge); therefore, to optimize the energy conversion efficiency of the modular power domain and reduce power loss, an efficiency optimization reward term is designed, and the characterization method is:

[0121] R efficiency =ΔV i (t)*I i (t); where, ΔV i (t) represents the voltage change of the i-th modular power domain at time step t; I i (t) represents the current of the i-th modular power domain at time step t;

[0122] In step S400, the corresponding action decision value is obtained through the state data, action decision, and reward function of the modular power domain. After the agent iterates repeatedly for several time steps, the action decision values form a global value network, and the state data and action decisions of all modular power domains form a global policy network. Specifically:

[0123] Step S401, obtaining the corresponding action decision value through the state data, action decision, and reward function of the modular power domain;

[0124] Step S402, repeating and iterating for several time steps;

[0125] Step S403, forming a global value network with the action decision values. The global value network is q(s i ,a i ,r i ,Φi); where, Φi represents the second network parameter of the local value network in the i-th modular power domain;

[0126] Step S404, forming a global policy network with the state data and action decisions of all modular power domains.

[0127] The parameters of the local policy network are represented as: s i ∈R ds ->a i ∈[-1,1] da ;

[0128] The parameters of the global policy network are represented as: s∈R N*ds ->g∈R N*da ; R represents the dimension identifier;

[0129] In step S500, the agent is updated in real time, and the states of different modular power domains are dynamically adjusted through the global policy network. Specifically:

[0130] Collect the first network parameters θ1, θ2,..., θN in the current local policy network of all agents, and calculate the global policy network parameter Θ through federated averaging, Θ = (θ1 + θ2 + … + θN) / N; realize the real-time update of the local policy network of the agent;

[0131] By calculating the value loss y, obtain the value loss gradient, and update the global value network according to the value loss gradient. The calculation of the value loss y is characterized as: y = ∑ t (r(t) + γV(s t+1 ) - V(s t )) 2 ; where γ represents the discount factor; V(s t ) represents the action decision value output by the global value network according to the state at time step t; r(t) represents the reward at time step t; V(s t+1 ) represents the action decision value output by the global value network according to the state at time step t + 1;

[0132] Usually, the expected value of state s is selected as the action decision value;

[0133] Dynamically adjust the supply current and heat dissipation power of each module power domain through the real-time updated global policy network;

[0134] Realize current adjustment through a DC-DC converter;

[0135] Realize heat dissipation power regulation by controlling the cooling fan through PWM.

[0136] An optimized management system for device battery consumption based on intelligent control includes an initial configuration module, a model construction module, an iterative training module, and an intelligent control module;

[0137] The initial configuration module is used to initialize and configure the modular power domain, and is respectively equipped with a current sensor, a voltage sensor, and a temperature sensor;

[0138] The model construction module is used to establish a reinforcement learning model based on multi-agent; including: regarding different modular power domains in the battery power supply as agents, and regarding the battery power supply as the global policy network, and the agent has an independent local policy network and local value network;

[0139] The iterative training module is used to divide time steps. At each time step, the agent respectively obtains the state data of the modular power domain and makes action decisions through an independent local policy network. Based on the change amount of the state data of the modular power domain after the action decision is executed, a reward function is designed. The corresponding action decision value is obtained through the state data of the modular power domain, the action decision, and the reward function. After the agent iteratively repeats several time steps, a global value network is formed with the action decision values, and a global policy network is formed with the state data and action decisions of all modular power domains.

[0140] The intelligent control module is used to update the agent in real time and dynamically adjust the states of different modular power domains through the global policy network.

[0141] Furthermore,

[0142] The intelligent control module includes a power supply current adjustment unit and a heat dissipation power control unit;

[0143] The power supply current adjustment unit is used to independently control the power switch and power supply current of the modular power domain;

[0144] The heat dissipation power control unit is used to control the power of the cooling system.

[0145] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claimed rights.

Claims

1. An optimized management method for device battery consumption based on intelligent control, characterized in that: The method for optimizing the management of device battery consumption specifically includes the following steps: Step S100: Based on the modular power domain, establish a reinforcement learning model based on multi-agent; including: regarding different modular power domains in the battery power supply as agents, regarding the battery power supply as a global policy network, and each agent has an independent local policy network and local value network; Step S200: Divide the time step. At each time step, each agent respectively obtains the state data of the modular power domain and makes an action decision through the independent local policy network; Step S300: Design a reward function based on the change amount of the state data of the modular power domain after executing the action decision; Step S400: Obtain the corresponding action decision value through the state data, action decision and reward function of the modular power domain. After the agent repeats and iterates for several time steps, form a global value network with the action decision value, and form a global policy network with the state data and action decisions of all modular power domains; Step S500: Update the agent in real time and dynamically adjust the states of different modular power domains through the global policy network.

2. The method for optimizing the management of device battery consumption based on intelligent control according to claim 1, wherein: In step S100, based on the modular power domain, establish a reinforcement learning model based on multi-agent; including: regarding different modular power domains in the battery power supply as agents, regarding the battery power supply as a global policy network, and each agent has an independent local policy network and local value network. Specifically: Regarding different modular power domains in the battery power supply as agents A i , where i = 1, 2, …, N, N represents the total number of modular power domains; i represents the quantity identifier of the modular power domain Each agent includes: Local policy network π θi (a i |s i ); The input is the state s i , and the output is the action a i ; θi represents the first network parameter; Local value network Q Φi (s i ,a i ); The local value network Q Φi (s i ,a i ) is used to evaluate the value of the state-action pair; Φi represents the second network parameter; Use the battery power supply as the global policy network Π Θ (g|s); where g represents the global action vector; s represents the concatenation of all agent states; Θ represents the global network parameters; The input of the global policy network is the global state vector, and the output is the coordination instruction.

3. The method for optimizing the management of device battery consumption based on intelligent control according to claim 2, characterized in that: Including: Define the state space: s i = [V i , I i , T i , SOC i , SOH i ; Among them, V i represents the voltage value; I i represents the current value; T i represents the temperature value; SOC i represents the state of charge; SOH i represents the state of health; Define the action space: a i = [ΔI i , B i ; Among them, ΔI i represents the supply current adjustment amount; B i represents the heat dissipation control coefficient.

4. The method for optimizing the management of device battery consumption based on intelligent control according to claim 3, characterized in that: In step S200, divide the time step. At each time step, each agent respectively obtains the state data of the modular power domain and makes an action decision through the independent local policy network. Specifically: Step S201: Set a fixed time step Δt, and the discrete time step sequence is t = k * Δt; where k represents the number of iterations, and k is a positive integer; make a decision at each time step t; Step S202: At time step t, each agent synchronously collects the state s i ; Step S203, the local policy network calculates the action, and the calculation formula is represented as: a i (t)=π θi (a i ∣s i )+ε t ; where a i (t) represents the action at time step t; ε t represents the exploration noise term; Step S204: Action execution, including: Calculate the current value of the next time step based on the current current value and the supply current adjustment amount; Calculate the actual heat dissipation power according to the heat dissipation control coefficient.

5. The method for optimizing and managing the battery consumption of a device based on intelligent control according to claim 4, characterized in that: In step S300, design a reward function based on the change amount of the state data of the modular power domain after executing the action decision. Specifically: The reward function is based on the change in state data Δs i =s i (t + 1)-s i (t) design; Agent A i Reward r at time step t i (t) is characterized as: r i (t) = w1 * R balance + w2 * R safety + w3 * R efficiency ; Among them, R balance represents the SOC balance reward term; R safety represents the temperature safety reward term; R efficiency represents the efficiency optimization reward term; w1, w2, and w3 are the corresponding reward weight coefficients respectively, satisfying w1 + w2 + w3 = 1.

6. The method for optimizing the management of device battery consumption based on intelligent control according to claim 5, characterized in that: In step S400, obtain the corresponding action decision value through the state data, action decision and reward function of the modular power domain. After the agent repeats and iterates for several time steps, form a global value network with the action decision value, and form a global policy network with the state data and action decisions of all modular power domains. Specifically: Step S401: Obtain the corresponding action decision value through the state data, action decision and reward function of the modular power domain; Step S402: Repeat and iterate for several time steps; Step S403: Construct a global value network with action decision values, where the global value network is q(s i , a i , r i , Φi); where Φi represents the second network parameter of the local value network in the ith modular power domain; Step S404: Form a global policy network with the state data and action decisions of all modular power domains.

7. The method for optimizing the management of device battery consumption based on intelligent control according to claim 6, characterized in that: In step S500, the agent is updated in real time, and the states of different modular power domains are dynamically adjusted through the global policy network. Specifically: Collect the first network parameters θ1, θ2, ..., θN in the current local policy network of all agents, and calculate the global policy network parameter Θ through the federated averaging method, Θ = (θ1 + θ2 + … + θN) / N; realize the real-time update of the local policy network of the agent; Obtain the value loss gradient by calculating the value loss, and update the global value network according to the value loss gradient; Dynamically adjust the supply current and heat dissipation power of each module power domain through the globally policy network updated in real time.

8. The method for optimizing the management of device battery consumption based on intelligent control according to claim 5, characterized in that: Specifically: Set the SOC balance reward term according to the magnitude of the deviation of the SOC change of the modular power domain from the global average; Adjust the cooling power according to the preset heat dissipation control coefficient, and design the temperature safety reward term.

9. The method for optimizing the management of device battery consumption based on intelligent control according to claim 7, characterized in that: The dynamic adjustment of the supply current and heat dissipation power of each module power domain through the globally policy network updated in real time includes: The supply current and heat dissipation power are independently controlled respectively.

Images