Intelligent energy storage control system and method for a battery of a power unit

Abstract

The application discloses a kind of intelligent energy storage control system and method of unit battery, the system includes: data acquisition module: the battery performance parameter and environmental parameter of battery pack are collected, battery performance parameter includes charging efficiency, discharge efficiency and battery resistance, environmental parameter includes temperature variation parameter, humidity variation parameter;Environment mapping module: an optimized and adjusted deep neural network is used to construct environment mapping model, battery performance parameter and environmental parameter are used as model input, and mapping influence coefficient is output;Intelligent distribution module: the output of environment mapping model and battery performance feedback data are used as the state and reward of reinforcement learning to construct reinforcement learning framework, and the selection strategy of reinforcement learning framework is optimized to obtain optimal strategy distribution scheme, and based on accurate battery state perception, energy storage control strategy that is more suitable for real-time state of battery and environmental condition can be formulated.

Description

Technical Field

[0001] This invention relates to the field of battery data processing technology, and in particular to an intelligent energy storage control system and method for generator battery. Background Technology

[0002] Currently, generator battery is widely used as an energy storage device in many scenarios, such as distributed energy systems and uninterruptible power supplies (UPS). However, with the growth of energy demand and the increasing requirements for energy storage efficiency and battery life, traditional generator battery energy storage control technology has gradually exposed many problems.

[0003] Existing energy storage control systems typically measure basic parameters such as battery charging and discharging efficiency and internal resistance directly for energy storage control. However, under different temperature conditions, the battery charging and discharging efficiency can fluctuate by up to 30%, and the internal resistance can change by 20%-50%. At the same time, humidity affects the chemical stability of the battery. In environments with high humidity, the battery's self-discharge rate will increase by 15%-25%.

[0004] On the other hand, traditional control systems lack intelligence and adaptability when adjusting battery energy storage strategies. They cannot dynamically optimize energy storage schemes based on real-time battery status and environmental changes. In practical applications, using fixed strategies for energy storage control without adjusting strategies based on battery status and environmental changes will lead to premature battery aging, reduced actual service life, and a 10%-15% decrease in energy storage efficiency. Therefore, an intelligent energy storage control system and method for generator batteries is proposed here. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above objectives, the present invention proposes the following technical solution:

[0006] An intelligent energy storage control system for generator battery includes:

[0007] Data acquisition module: Collects battery performance parameters and environmental parameters of the battery pack;

[0008] Among them, the battery performance parameters include at least charging efficiency, discharging efficiency and battery internal resistance, and the environmental parameters include at least temperature change parameters and humidity change parameters.

[0009] Environment mapping module: Constructs an environment mapping model through an optimized and adjusted deep neural network, taking battery performance parameters and environmental parameters as input data for the model, and outputting mapping influence coefficients;

[0010] Intelligent allocation module: Based on the output of the environment mapping model and battery performance feedback data, the module uses the state and reward of reinforcement learning to construct a reinforcement learning framework and optimize the selection strategy of the reinforcement learning framework to obtain the optimal policy allocation scheme.

[0011] Among them, the optimized deep neural network dynamically adjusts the number of hidden layer nodes by acquiring the joint entropy of the model input data; the selection strategy of the optimized reinforcement learning framework is implemented by combining the ∈-greedy strategy with the recurrent neural network.

[0012] The charging efficiency is obtained by measuring the input energy and effective stored energy during the battery charging process; the discharging efficiency is obtained by measuring the output energy and discharge energy during the discharge time using a power sensor; the battery internal resistance is obtained by injecting a small AC signal of known frequency and amplitude into the battery using an internal resistance measuring sensor and measuring the response voltage across the battery terminals.

[0013] The temperature change parameter is obtained by converting temperature changes into voltage signals through thermocouple temperature sensors, which are distributed and installed at different locations in the battery pack. The voltage value is measured and combined with the calibration curve. The humidity change parameter is obtained by measuring the capacitance value as the ambient humidity changes through a capacitive humidity sensor placed in the environment where the battery pack is located. The value is then converted by a signal processing circuit.

[0014] The process of constructing the environment mapping model is as follows:

[0015] The environment mapping model includes an input layer, an optimized hidden layer, and an output layer;

[0016] The number of input layer nodes is determined based on the types of environmental parameters and battery performance parameters collected;

[0017] The number of hidden layer nodes after optimization is set to twice the number of input layer nodes, and a 3-layer hidden layer structure is adopted. Each hidden layer uses the ReLU activation function and is adjusted dynamically through a joint entropy adjustment strategy.

[0018] The number of output layer nodes is determined based on the type of predicted battery mapping data and linear activation is used.

[0019] The joint entropy dynamic adjustment strategy includes:

[0020] The model input data is divided into data groups according to time windows;

[0021] Calculate the joint entropy H(z) for each set of data, where z represents the joint data, z = [η charge η discharge [,R,T,H];

[0022] When the joint entropy H(z) > H upper At that time, the hidden layer is increased by 10% of the current number of nodes;

[0023] When the joint entropy H(z) <H lower When the number of nodes is reduced, the hidden layer is reduced by 5% of the current number of nodes;

[0024] When the joint entropy H lower ≤H(z)≤H upper At the same time, the number of hidden layer nodes remains unchanged.

[0025] The battery performance feedback data includes at least the actual charging efficiency, the actual discharging efficiency, and the battery life change rate.

[0026] The reinforcement learning framework uses the output of the environment mapping model and battery performance feedback data as the state and reward signals in the reinforcement learning process.

[0027] The reinforcement learning framework includes an agent, an environment V, and a state S;

[0028] The intelligent agent controls the environment mapping model and weight allocation; the environment V includes at least the charging and discharging process of the battery at different times, the physical environment, and the environment mapping model itself; the state S is composed of the output of the environment mapping model and battery performance feedback data.

[0029] The agent selects action A based on state S using an ∈-greedy strategy combined with a recurrent neural network. It then comprehensively evaluates the existing optimal action, along with interference actions from the recurrent neural network with a specific probability, and compares them with actions in the preliminary optimal action set to ultimately determine the action to be executed.

[0030] The process of selecting actions using the ∈-greedy strategy combined with a recurrent neural network is as follows:

[0031] Create and initialize a table storing the expected cumulative reward Q(S, A) for performing different actions A in different states S, and input the current state S into a recurrent neural network. t Output a series of candidate interference actions A adversarial ;

[0032] At time step t, the output data of the environment mapping model and the battery performance feedback data are combined to form state S. t ;

[0033] The agent queries the Q-table to find the current state S. t Given actions with specified Q values, form a preliminary optimal action set A. optimal ;

[0034] The set of perturbation actions A provided by the recurrent neural network adversarial In the game, one action is randomly selected from a uniform distribution as the antagonistic action. Selected confrontation action With the initial optimal action set A optimal The actions in the process are evaluated uniformly;

[0035] Based on the evaluation results, the action with the highest score will be selected as the final action to be performed.

[0036] A method for intelligent energy storage control of generator battery, comprising the following steps:

[0037] S1: Collect battery performance parameters and environmental parameters of the battery pack. Battery performance parameters include charging efficiency, discharging efficiency and battery internal resistance. Environmental parameters include temperature change parameters and humidity change parameters.

[0038] S2: An environment mapping model is constructed through an optimized and adjusted deep neural network. Battery performance parameters and environmental parameters are used as model inputs, and the mapping influence coefficients are output.

[0039] S3: Based on the output of the environment mapping model and the battery performance feedback data, a reinforcement learning framework is constructed using the state and reward of the reinforcement learning framework, and the selection strategy of the reinforcement learning framework is optimized to obtain the optimal policy allocation scheme.

[0040] The present invention has the following beneficial effects:

[0041] In this invention, firstly, the number of hidden layer nodes is dynamically adjusted based on the joint entropy of the input data through the optimized deep neural network in the environment mapping module. This enables more accurate capture of the complex mapping relationship between environmental parameters (temperature, humidity) and battery performance parameters (charging efficiency, discharging efficiency, internal resistance). The mapping influence coefficient output by the environment mapping model clearly reflects the significant impact of temperature on the overall battery performance, which helps to more accurately predict changes in battery performance and improve the system's ability to perceive battery status.

[0042] Secondly, by combining the ∈-greedy strategy in the intelligent allocation module with a recurrent neural network, the agent can make full use of the time-series characteristics of historical battery charging and discharging data and environmental parameters. Based on existing experience, the agent can select the optimal action. At the same time, it can explore new strategies by accepting the interference actions given by the recurrent neural network with a specific probability, which increases the diversity of action selection. After multiple iterations and optimizations, it can better obtain the optimal strategy allocation scheme under different battery states and environments, effectively balancing the charging and discharging efficiency and lifespan of the battery, and improving the overall performance of the energy storage system.

[0043] Finally, the environment mapping module captures the mapping relationship between the environment and battery performance parameters, providing an accurate basis for energy storage control strategies. The intelligent allocation module obtains the optimal strategy allocation scheme by combining the ∈-greedy strategy of the recurrent neural network. The combination of the two can formulate energy storage control strategies that are more in line with the real-time state of the battery and environmental conditions based on accurate battery state perception. It can accurately adjust the charging power and discharging strategy according to the significant impact of the environment on battery performance and the characteristics of historical charging and discharging data, so that the battery can maintain efficient and stable operation. Attached Figure Description

[0044] Figure 1 This is a system block diagram of an intelligent energy storage control system and method for generator battery proposed in this invention.

[0045] Figure 2 This is a flowchart illustrating the steps of an intelligent energy storage control system and method for generator battery proposed in this invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Example 1

[0048] like Figure 1 As shown, the present invention proposes an intelligent energy storage control system for generator battery, the system comprising:

[0049] Data acquisition module: Collects battery performance parameters and environmental parameters of the battery pack. Battery performance parameters include charging efficiency, discharging efficiency and battery internal resistance. Environmental parameters include temperature change parameters and humidity change parameters.

[0050] The process for obtaining battery performance parameters is as follows:

[0051] Charging efficiency sensor: Calculates charging efficiency η by measuring the input energy and effective stored energy during battery charging. charge ;

[0052] Discharge efficiency sensor: During battery discharge, a power sensor measures the output power of the discharge circuit and the battery's discharge power. The discharge efficiency η is calculated from the output energy and discharge energy over the discharge time. discharge ;

[0053] Battery internal resistance measurement sensor: This internal resistance measurement sensor uses the AC injection method. It injects a small AC signal with a known frequency and amplitude into the battery, measures the response voltage across the battery terminals, and calculates the battery internal resistance R based on Ohm's law.

[0054] The process of obtaining environmental parameters is as follows:

[0055] High-precision thermocouple temperature sensors are selected and installed at different locations in the battery pack to obtain the temperature distribution of the battery pack. The thermocouples are based on the thermoelectric effect and convert temperature changes into voltage signals. By measuring the voltage value and combining it with the calibration curve, the corresponding temperature value T is obtained.

[0056] Humidity sensor: A capacitive humidity sensor is used and placed in the environment where the battery pack is located to monitor the ambient humidity in real time. The capacitance value of the capacitive humidity sensor will change with the change of ambient humidity. The capacitance value is measured and converted into a humidity value H by the signal processing circuit.

[0057] The ARM-based microcontroller serves as the core of the data acquisition module, converting the analog signals output by the sensors into digital signals and transmitting the digital signals to the environment mapping module.

[0058] Environment mapping module: Constructs an environment mapping model through an optimized and adjusted deep neural network, taking battery performance parameters and environmental parameters as model inputs, and outputting mapping influence coefficients;

[0059] The optimized and adjusted deep neural network is based on the hidden layer of the original neural network, and the hidden layer nodes are adjusted according to the joint entropy of battery performance parameters and environmental parameters.

[0060] The process of constructing the environment mapping model is as follows:

[0061] The environment mapping model is based on a deep neural network with an optimized hidden layer node.

[0062] Deep neural networks consist of an input layer, hidden layers, and an output layer;

[0063] The environment mapping model includes an input layer, an optimized and adjusted hidden layer, and an output layer.

[0064] Constructing the input layer of the environment mapping model: The number of nodes in the input layer is determined based on the types of collected environmental parameters and battery performance parameters. One type corresponds to one node. Let the number of nodes in the input layer be y.

[0065] Constructing the hidden layers of the environment mapping model: The initial number of hidden layer nodes is set to twice the number of input layer nodes, i.e., the initial number of hidden layer nodes is 2y. A multi-layer hidden layer structure is adopted, with 3 hidden layers. Each hidden layer uses the ReLU activation function as follows:

[0066] h j =ReLU(W j x i +b j )

[0067] Where j represents the level index (j = 1, 2, 3), h j W is the output of the j-th hidden layer. j Let b be the weight of the j-th hidden layer. j For the bias term of the j-th hidden layer, x i This refers to the data value of the i-th input.

[0068] The process of optimizing and adjusting hidden layer nodes is as follows:

[0069] Group the input data according to certain features or dimensions;

[0070] For example, the input data, consisting of battery performance parameters (charging efficiency, discharging efficiency, battery internal resistance) and environmental parameters (temperature, humidity), is divided into k data windows according to the time series, and each window contains a certain number of sample data.

[0071] For each group, calculate the frequency of each data value and use it as an estimate of the probability of that data value occurring. Suppose there are N samples in a data group, where the i-th data value is x. i The number of times it appears is n i x i The probability of occurrence is

[0072] Calculate the joint entropy for each group of data:

[0073]

[0074] Where z represents joint data, z = [η charge η discharge [,R,T,H],p z It represents the probability of the joint data appearing simultaneously;

[0075] By calculating p z (The probability of multiple parameters occurring simultaneously), joint entropy can reveal hidden relationships between parameters, such as the distribution pattern of charging efficiency under specific humidity and temperature conditions, help discover the influence mechanism of multi-parameter synergy on battery performance, provide guidance for deep neural networks to capture complex correlation features, and optimize the network's ability to learn multi-parameter coupling patterns.

[0076] Based on the calculated joint entropy, an optimization strategy for adjusting the number of hidden layer nodes is formulated.

[0077] Setting a threshold: By setting an upper threshold H for the joint entropy.upper and lower threshold H lower These two thresholds are used to define different levels of data complexity;

[0078] For example, when the joint entropy H(z) > H upper At that time, it was considered that the data complexity was high;

[0079] When the joint entropy H(z) <H lower At that time, the data complexity was considered to be low;

[0080] Upper threshold H lower ≤H(z)≤H upper At that time, the data complexity was considered to be within a moderate range;

[0081] When the joint entropy of the input data H(z) > H upper That is, when the data complexity is large, the number of hidden layer nodes is increased by 10% of the current number of hidden layer nodes;

[0082] When the joint entropy H(z) of the input data <H lower That is, when the data complexity is low, reduce the number of hidden layer nodes. Too many nodes can easily lead to overfitting on simple data. Reducing the number of nodes can make the model more concise and effective. The number of nodes should be reduced by 5% of the current number of nodes.

[0083] When H lower ≤H(z)≤H upper That is, when the data complexity is moderate, the number of hidden layer nodes remains unchanged;

[0084] Construct the environment mapping model output layer:

[0085] The number of output layer nodes is determined based on the type of predicted battery mapping data;

[0086] For example, if we want to predict three battery performance parameters, namely charging efficiency, discharging efficiency, and battery internal resistance, then the number of output layer nodes is set to 3, because each parameter to be predicted corresponds to a node in the output layer, which is used to output the predicted value of that parameter.

[0087] Design the output layer structure: The output layer uses linear activation;

[0088] Specifically, the output layer predicts the relationship between the environment and battery performance, so the number of nodes in the output layer is set to 2. Let the weight matrix from the third hidden layer to the output layer be W4, and the bias vector be b4. Then the output of the third hidden layer is h3, and the output of the output layer is represented as... Because battery performance parameters do not have a range of values ​​in a practical sense and do not need to be output within a specific interval, linear activation of battery performance parameters can directly output predicted values, mapping the features extracted by the hidden layer to the specific battery mapping data predicted values.

[0089] The mean square error is used as the loss function to measure the gap between the output result of the model and the mapping relationship between the actual environment and battery performance. Let the quantified value of the mapping relationship between the actual environment and battery performance be Y, then the loss function

[0090]

[0091] where N is the number of samples in the data grouping;

[0092] Then the output result of the environment mapping model is F1 represents the mapping degree value between the current environment (temperature) and the overall battery performance (charge-discharge efficiency, battery internal resistance) (one-to-many). The larger F1 is, the more significant the impact of the current environmental parameter (temperature) on the overall performance of the battery. The larger F2 is, the more significant the impact of the current environmental parameter (humidity) on the overall performance of the battery; <o

[0093] Example: In a battery environment of a unit:

[0094] Environmental parameters: temperature is 15°C - 30°C, humidity is 30% - 60%;

[0095] Battery performance parameters: charge efficiency is 80% - 90%, discharge efficiency is 75% - 85%, battery internal resistance is 10mΩ - 20mΩ;

[0096] Because there are 5 parameters, the number of nodes in the input layer of the environment mapping model is set to 5. The normalized data is composed into an input vector and input into the model;

[0097] The number of nodes in the initial hidden layer is set to twice the number of nodes in the input layer, that is, 10 nodes, and 3 hidden layers are constructed;

[0098] The input data within one week is divided into windows with 12 samples each, and a total of 14 data windows are obtained. Taking one data window as an example, according to the joint entropy formula, the joint entropy H of the joint data within this window is calculated as H = 1.8. Set the upper threshold H upper = 2.0, and the lower threshold H lower = 1.2. Since 1.2 < H = 1.8 < 2.0, it indicates that the complexity of this grouped data is moderate, and the number of nodes in the hidden layer remains unchanged at 10;

[0099] The environment mapping model outputs values through the output layer

[0100] Interpretation of results: The first value of 2.5 is relatively large, indicating that the current environmental parameter (temperature) has a significant impact on the overall battery performance. The second value of 1.2 is relatively small, indicating that in another dimension, the correlation between the environment (humidity) and battery performance (such as the effect of humidity on battery internal resistance) is relatively weak. Based on these results, in the actual intelligent energy storage control system of the generator battery, more attention should be paid to the impact of temperature on battery performance to avoid battery overheating affecting performance and lifespan.

[0101] Intelligent allocation module: Based on the output of the environment mapping model and battery performance feedback data as the state and reward of reinforcement learning, a reinforcement learning framework is constructed, and the selection strategy of the reinforcement learning framework is optimized to obtain the optimal policy allocation scheme;

[0102] Battery performance feedback data includes actual charging efficiency, actual discharging efficiency, and battery life change rate.

[0103] A reinforcement learning framework is constructed, which uses the model's output (mapped influence coefficient) and battery performance feedback data as the state and reward signals for reinforcement learning. Different output strategies are continuously tried so that the prediction results gradually approach the optimal solution.

[0104] Reinforcement learning frameworks include:

[0105] Intelligent agent: controls the environment mapping model and data allocation;

[0106] Environment V: The environment includes the actual operating status of the battery, which consists of the battery's charging and discharging process at different times, the physical environment (temperature, humidity, etc.), and the environment mapping model itself. The environment will change accordingly based on the actions of the intelligent agent and provide feedback information.

[0107] State S: The state is composed of the output of the environment mapping model and battery performance feedback data. The output of the environment mapping model is...

[0108] Actual battery performance feedback data is as follows: in, This indicates the actual charging efficiency. η represents the actual discharge efficiency. * This indicates the rate of change in battery life, which is obtained directly through the battery management system (BMS) or related monitoring equipment.

[0109] Based on the current state S, the agent selects action A using an improved ∈-greedy strategy (combined with a recurrent neural network). This involves considering the optimal action based on prior experience, and accepting perturbation actions from the recurrent neural network with a specific probability. For example, the agent first determines a preliminary set of optimal actions, then randomly selects one from the actions provided by the recurrent neural network, and comprehensively evaluates it against the actions in the preliminary optimal action set to ultimately determine the action to be executed.

[0110] The process of combining the ∈-greedy strategy of a recurrent neural network is as follows:

[0111] Create a table to store the expected cumulative reward Q(S, A) for performing different actions A in different states S. In the initial stage, initialize all Q values ​​to a small random value, such as 0.

[0112] Example: Create a table Q(S, A). Suppose that state S includes two types, S1 and S2, and action A includes two types, A1 (adjust charging power) and A2 (adjust discharge strategy). Initially, Q(S1, A1) = 0, (S1, A2) = 0, Q(S2, A1) = 0, and Q(S2, A2) = 0.

[0113] The recurrent neural network takes the current state S as input. t Where t is the time step, S t This represents the state at time step t, and the output is a series of candidate perturbation actions. At the same time, the weights and biases of the recurrent neural network are initialized.

[0114] Specifically, recurrent neural networks excel at processing sequential data, capturing long-term dependencies in time series through their memory units. For example, they can analyze historical battery charge and discharge data and the changing patterns of environmental parameters over time. When combined with recurrent neural networks, they can be used to process state inputs (such as time series data of battery performance) and extract sequence features. The ∈-greedy strategy, on the other hand, determines the action to explore or utilize based on the state processed by the recurrent neural network.

[0115] Probability parameter: Sets the specific probability of accepting actions from the recurrent neural network. And the ∈ value in the traditional ∈-greedy strategy;

[0116] At time step t, the agent collects relevant information about the current environment, including the output of the environment mapping model and battery performance feedback data, and integrates them into state S. t S t Includes information on temperature, humidity, predicted charging efficiency, actual discharging efficiency, and battery internal resistance.

[0117] Based on the Q-list, a preliminary optimal set of actions is determined. The agent then queries the Q-list to find the action that is optimal in the current state S.t The actions with higher Q values ​​are used to form a preliminary optimal action set A. optimal ;

[0118] For example: If the current state S t Q(S) corresponding to action A1 t A1) = 0.2, Q(S) of action A2 t If A2) = 0.2, then the initial optimal action set A optimal =A1;

[0119] Specifically, these actions are based on the agent's existing experience and are considered to yield a good expected cumulative reward in the current state;

[0120] Set the current state S t The input is fed into a recurrent neural network, which then generates a series of perturbation actions A. adversarial The purpose of these actions is to mislead the agent into choosing suboptimal actions, thereby increasing the diversity and exploratory nature of action choices. For example, the perturbation actions output by recurrent neural networks include parameter adjustments or policy changes that differ from the agent's usual experience.

[0121] Specifically, perform action A. t Afterwards, environmental feedback reward r t =0.3, transferred to the new state S t+1 Update the Q value using the Q-learning formula:

[0122] Q(S t A t )=Q(S t A t )+αr t +γmax A Q(S t+1 A t ),AQ(S t A t )

[0123] Here, we assume α is the learning rate and it is 0.2, γ is the discount factor and it is γ = 0.9, and the new state S t+1 max A Q(S t+1 A t If ) = 0.4, then after the update:

[0124] Q(S t , A1)=0.2+0.2×[0.3+0.9×0.4-0.2]=0.292;

[0125] Then, from the set of disturbance actions A provided by the recurrent neural network... adversarialIn the process, one action is randomly selected as the antagonistic action according to a uniform distribution;

[0126] Selected confrontation action With the initial optimal action set A optimal The actions in the process are evaluated together;

[0127] Specifically, the evaluation method involves querying the Q-table again to obtain the Q-value for each action;

[0128] Based on the evaluation results, the action with the highest score will be selected as the final action to be performed.

[0129] For example, in a comprehensive assessment, the confrontational actions If the score is higher than all actions in the initial optimal action set, then Otherwise, select the action with the highest score from the initial optimal action set as...

[0130] By iterating multiple times (e.g., 1000 disturbance actions), the Q-table is continuously updated. When the Q-value of the disturbance action converges (the change in Q-value is less than 0.01 after 50 consecutive iterations), the combination of actions that maximizes Q(S,A) in each state in the Q-table is the optimal policy allocation scheme.

[0131] Finally, A1 is selected in state S1 and A2 is selected in state S2, forming the optimal control strategy for different battery states and environments;

[0132] For example: State S1 (high temperature, high SOC interference): Select A1 (reduce charging power interference by 50%) to reduce battery loss caused by high temperature and high load.

[0133] State S2 (normal temperature, moderate SOC interference): Select A2 (maintain normal charging power) to improve charging efficiency while ensuring battery life.

[0134] Example 2

[0135] like Figure 2 As shown, a smart energy storage control method for generator battery is described, and the steps of the method are as follows:

[0136] S1: Collect battery performance parameters and environmental parameters of the battery pack. Battery performance parameters include charging efficiency, discharging efficiency and battery internal resistance. Environmental parameters include temperature change parameters and humidity change parameters.

[0137] S2: An environment mapping model is constructed through an optimized and adjusted deep neural network. Battery performance parameters and environmental parameters are used as model inputs, and the mapping influence coefficients are output.

[0138] S3: Based on the output of the environment mapping model and the battery performance feedback data, a reinforcement learning framework is constructed using the state and reward of the reinforcement learning framework, and the selection strategy of the reinforcement learning framework is optimized to obtain the optimal policy allocation scheme.

[0139] In the application, several formulas are calculated by removing dimensions and taking their numerical values. The formulas are established by collecting a large amount of data and simulating the most recent real situation. Some coefficients or weights in the formulas are set by those skilled in the art according to the actual situation, so they will not be elaborated here.

[0140] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0141] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent energy storage control system for generator battery, characterized in that, include: Data acquisition module: Collects battery performance parameters and environmental parameters of the battery pack. Battery performance parameters include charging efficiency, discharging efficiency and battery internal resistance. Environmental parameters include temperature change parameters and humidity change parameters. Environment mapping module: Constructs an environment mapping model through an optimized and adjusted deep neural network, taking battery performance parameters and environmental parameters as model inputs, and outputting mapping influence coefficients; Intelligent allocation module: Based on the output of the environment mapping model and battery performance feedback data as the state and reward of reinforcement learning, a reinforcement learning framework is constructed, and the selection strategy of the reinforcement learning framework is optimized to obtain the optimal policy allocation scheme; The basic architecture of the environment mapping model is an optimized and adjusted deep neural network; The optimized deep neural network dynamically adjusts the number of hidden layer nodes by acquiring the joint entropy of the input data; Optimize the selection strategy of reinforcement learning framework through The strategy is implemented using a recurrent neural network; Joint entropy dynamic adjustment strategies include: The input parameters are divided into data groups according to the time window; Calculate the joint entropy of each set of data. , Indicates joint data, ; When joint entropy When the number of nodes is increased, a hidden layer is added at a rate of 10% of the current number of nodes. When joint entropy At any time, the hidden layer is reduced by 5% of the current number of nodes; When joint entropy At the same time, keep the number of hidden layer nodes unchanged; The optimized reinforcement learning framework uses the output of the environment mapping model and battery performance feedback data as the state and reward signals for reinforcement learning. The reinforcement learning framework includes an agent and an environment. and state ; The intelligent agent controls the environment mapping model and weight allocation; The environment This includes the battery's charging and discharging process at different times, its physical environment, and the environmental mapping model itself. The state It consists of the output of the environmental mapping model and battery performance feedback data.

2. The intelligent energy storage control system for a generator battery according to claim 1, characterized in that, The charging efficiency is obtained by measuring the input energy and effective stored energy during the battery charging process; The discharge efficiency is obtained by measuring the output energy and discharge energy during the discharge time using a power sensor. The battery's internal resistance is obtained by injecting a small AC signal of known frequency and amplitude into the battery through an internal resistance measurement sensor and measuring the response voltage across the battery terminals. The temperature change parameters are obtained by converting temperature changes into voltage signals through thermocouple temperature sensors, which are distributed and installed at different locations in the battery pack. The voltage values ​​are measured and combined with the calibration curve. The humidity change parameter is obtained by measuring the change in capacitance value as the ambient humidity changes through a capacitive humidity sensor placed in the environment where the battery pack is located, and then converting the result through a signal processing circuit.

3. The intelligent energy storage control system for a generator battery according to claim 1, characterized in that, The process of constructing the environment mapping model is as follows: The environment mapping model includes an input layer, an optimized hidden layer, and an output layer; The number of input layer nodes is determined based on the types of environmental parameters and battery performance parameters collected; The initial number of hidden layer nodes is set to the number of input layer nodes. times, and adopt Hidden layer structure, each hidden layer uses The activation function is adjusted using a joint entropy dynamic adjustment strategy. The number of output layer nodes is determined based on the type of predicted battery mapping data and linear activation is used.

4. The intelligent energy storage control system for a generator battery according to claim 1, characterized in that, The battery performance feedback data includes actual charging efficiency, actual discharging efficiency, and battery life change rate.

5. The intelligent energy storage control system for a generator battery according to claim 1, characterized in that, The intelligent agent is based on the state ,use The strategy combines recurrent neural networks to select actions. Based on the optimal action derived from existing experience, and the interference actions given by the recurrent neural network with a specific probability, the actions in the preliminary optimal action set are comprehensively evaluated to finally determine the action to be executed. .

6. The intelligent energy storage control system for a generator battery according to claim 5, characterized in that, The The process of selecting actions using a strategy combined with a recurrent neural network is as follows: Create and store in different states Perform different actions Expected cumulative rewards The table is initialized, and the current state is input through a recurrent neural network. Output a series of candidate interference actions ; At time step The output of the environmental mapping model and the battery performance feedback data will be combined with the synthetic state. ; Based on agent query - Table, find the current state The following is specified The actions that determine the value form a preliminary set of optimal actions. ; The set of perturbation actions provided by the recurrent neural network In the game, one action is randomly selected from a uniform distribution as the antagonistic action. Selected confrontation actions With the initial optimal action set The actions in the process are evaluated together; Based on the evaluation results, the action with the highest score will be selected as the final action to be performed. .

7. A method for intelligent energy storage control of a generator battery, using the system described in any one of claims 1 to 6, characterized in that, The steps are as follows: S1: Collect battery performance parameters and environmental parameters of the battery pack. Battery performance parameters include charging efficiency, discharging efficiency and battery internal resistance. Environmental parameters include temperature change parameters and humidity change parameters. S2: An environment mapping model is constructed through an optimized and adjusted deep neural network. Battery performance parameters and environmental parameters are used as model inputs, and the mapping influence coefficients are output. S3: Based on the output of the environment mapping model and the battery performance feedback data, a reinforcement learning framework is constructed using the state and reward of the reinforcement learning framework, and the selection strategy of the reinforcement learning framework is optimized to obtain the optimal policy allocation scheme.

Images