Power source balancing method based on AI

Abstract

The invention discloses a power source balancing method based on AI, and belongs to the technical field of signal processing. The objective of the invention is to solve the technical problem of poor short-time stability and long-term stability balance effect of power source output in the prior art. The thermal runaway risk of the switch tube can be effectively warned in advance by fusing temperature and impedance angle information; the prediction error of voltage fluctuation can be effectively reduced through the LSTM sub-network, the PID parameter adjustment time can be shortened through gradient descent optimization, and the 24-hour drift rate can be effectively controlled; the output PWM duty ratio correction value is injected into the power source driving module through the high-speed digital interface, so that the time sequence consistency of the PWM correction value and voltage sampling can be ensured, and the ripple coefficient measurement error is effectively reduced; and the output quality of the power source can be reflected more comprehensively and reliably by monitoring the high-frequency ripples with short-term stability and the long-term drifting with long-term stability at the same time.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and more specifically to an AI-based power source balancing method. Background Technology

[0002] Existing high-precision three-phase power sources mainly adopt multi-phase power supply design, and their balancing scheme aims to achieve uniform distribution of loads in each phase and stable output.

[0003] Current high-precision three-phase power source balancing solutions mainly rely on technologies such as optimizing power load distribution, using reactive power compensation devices, and installing phase-to-phase balancing devices. Although these technologies are relatively mature, the following problems still exist in practical applications: the continuous changes in power load (such as temporary power consumption and seasonal power consumption) make it difficult to maintain a balanced load distribution in the long term, posing a challenge to dynamic load changes; if there is a lack of regular detection and adjustment of three-phase loads in distribution network management, imbalance problems may accumulate, affecting the long-term stability of the system. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-based power source balancing method to solve the technical problem of poor short-term and long-term stability balancing of power source output in existing solutions.

[0005] The objective of this invention can be achieved through the following technical solutions: An AI-based power source balancing method includes: A multi-mode measurement component is deployed between the output end and the load end of a high-precision three-phase power source to synchronously acquire the instantaneous values ​​of three-phase voltage, the effective values ​​of current, the junction temperature of the switching transistor, and the load impedance angle, generating a four-dimensional time-series data sequence. The four-dimensional time-series data sequence is preprocessed, and the spatiotemporal correlation is fused through an attention mechanism to construct a dynamic feature matrix. The generated dynamic feature matrix is ​​input into a pre-trained deep reinforcement learning model, and the output is a dual-time-scaled control strategy containing short-time voltage compensation and long-term parameter adjustment. The short-time voltage compensation is generated by predicting future voltage fluctuations through an LSTM sub-network to generate a PWM duty cycle correction value. The long-term parameter adjustment is based on the temperature drift trend in the feature matrix and optimizes the proportional coefficient and integral time of the PID controller through a gradient descent algorithm. The output PWM duty cycle correction value is injected into the power source drive module, and the compensated output voltage ripple coefficient and drift rate are collected synchronously. The voltage ripple coefficient and drift rate are analyzed to determine whether the high power source output is unbalanced. If unbalanced, the parameter optimization process is re-executed until the analysis result shows that the high power source output is balanced.

[0006] Preferably, a measurement component consisting of a voltage acquisition module, a temperature acquisition module, and an impedance angle acquisition module is deployed in series between the output terminal and the load terminal of the high-precision three-phase power source. The collected instantaneous values ​​of three-phase voltage, effective values ​​of three-phase current, surface temperatures of six switching transistors, and load impedance angles are aligned according to timestamps to form a four-dimensional time-series data sequence.

[0007] Preferably, when fusing spatiotemporal correlations through an attention mechanism, the mutual information entropy of voltage transient features and temperature drift features is calculated; the mutual information entropy is analyzed and dynamically assigned weight coefficients; When constructing the dynamic feature matrix, the extracted 16-dimensional high-frequency transient features, 6-dimensional temperature drift features, and 1-dimensional impedance angle features are weighted and summed to form the dynamic feature matrix.

[0008] Preferably, the pre-trained deep reinforcement learning model adopts the DDPG framework, which includes an Actor network and a Critic network. The Actor network corresponds to the policy network, and the Critic network corresponds to the value network. Both of them adopt a fully connected neural network structure.

[0009] Preferably, based on the predicted voltage fluctuation value The PWM duty cycle correction value is generated by calculation, and the relevant expression is as follows: ;in, This is the proportionality coefficient; is the integration coefficient, with a default value of 0.01; t is the integration time window.

[0010] Preferably, when generating long-term parameter adjustments through gradient descent optimization, the output voltage drift rate δU over 24 hours is used as the optimization objective, and the loss function is defined as follows: ;in, , where are the PID parameters to be optimized. This is the proportionality coefficient. The integration time; These are the proportional coefficient regularization coefficient and the integral time regularization coefficient, respectively. The output voltage drift rate over 24 hours. , These are the rated output voltage, the maximum voltage value within 24 hours, and the minimum voltage value, respectively. It is the square of the L2 norm.

[0011] Preferably, the Adam optimizer is used to update the PID parameters, and the relevant expression is: ;in, The PID parameters are for the j-th iteration; Here are the updated parameters for the (j+1)th iteration; j is the iteration index. For the loss function in gradient at; This is the learning rate.

[0012] Preferably, the output PWM duty cycle correction value is injected into the power source drive module through a high-speed digital interface; during injection, a hardware synchronization signal is used to ensure that the correction value is strictly synchronized with the power source switching cycle, so as to avoid voltage fluctuations caused by timing deviations.

[0013] Preferably, when collecting drift rate data for 24 hours, one effective voltage value is collected every 100ms, and the data is continuously collected for 24 hours to generate a voltage time sequence; the voltage ripple coefficient is calculated based on the peak value of the AC ripple component in the output voltage and the rated voltage.

[0014] Preferably, the calculated voltage ripple coefficient and drift rate are compared with the corresponding first balance threshold and second balance threshold, respectively. If the voltage ripple coefficient is less than or equal to the first balance threshold and the absolute value of the drift rate is less than or equal to the second balance threshold, then the high power source output is considered balanced. If the voltage ripple coefficient is greater than the first balance threshold and / or the drift rate is greater than the second balance threshold, the high power source output is determined to be unbalanced and an interrupt signal is generated. The parameter optimization process is re-executed through the interrupt signal to generate a new PWM duty cycle correction value, and the compensation and determination are repeated.

[0015] Compared to existing solutions, the beneficial effects achieved by this invention are: This invention, by fusing temperature and impedance angle information, can effectively provide early warning of the risk of thermal runaway of switching transistors compared to traditional single-physical-quantity acquisition; and by using an attention mechanism, it can effectively improve the signal-to-noise ratio of the feature matrix and reduce the output delay of the control strategy of the subsequent AI model.

[0016] This invention can effectively reduce the prediction error of voltage fluctuations through LSTM subnetwork, and can compensate for ripple in advance, thereby reducing the output voltage ripple coefficient; gradient descent optimization can shorten the PID parameter adjustment time and effectively control the 24-hour drift rate; the DDPG model has low output delay in the control strategy during load changes, and the system recovery time can be effectively reduced compared with PI control, thus enhancing robustness.

[0017] This invention ensures the timing consistency between the PWM correction value and voltage sampling by injecting the output PWM duty cycle correction value into the power source drive module through a high-speed digital interface, effectively reducing the measurement error of the ripple coefficient. By simultaneously monitoring the high-frequency ripple of short-term stability and the long-term drift of long-term stability, it can more comprehensively and reliably reflect the output quality of the power source compared to existing single indicators. By analyzing the voltage ripple coefficient and drift rate, it determines whether the high-power source output is balanced and dynamically triggers the re-execution of the parameter optimization process, realizing automated closed-loop parameter tuning, adapting to fast dynamic load scenarios, and further improving the reliability and adaptability of the balance. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings.

[0019] Figure 1 This is a flowchart of an AI-based power source balancing method according to the present invention. Detailed Implementation

[0020] 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.

[0021] like Figure 1 As shown, this invention is an AI-based power source balancing method, comprising: A multi-mode measurement component is deployed between the output and load terminals of a high-precision three-phase power source to synchronously acquire instantaneous three-phase voltage values, effective current values, switch junction temperature, and load impedance angle, generating a four-dimensional time-series data sequence. This four-dimensional time-series data sequence is preprocessed, and spatiotemporal correlations are fused using an attention mechanism to construct a dynamic feature matrix. The specific steps include: A measurement component consisting of a voltage acquisition module, a temperature acquisition module, and an impedance angle acquisition module is deployed in series between the output terminal and the load terminal of a high-precision three-phase power source. Among them, the voltage acquisition module consists of one Hall voltage sensor connected in series in each phase line. The sensor output signal is converted by a 16-bit ADC and then fed into the FPGA preprocessing unit. Current acquisition module: Each phase line is equipped with a Rogowski coil current sensor, and the output is connected to an integrator and then to a synchronous sampling circuit; Temperature acquisition module: Distributed fiber optic temperature sensors are attached to the heat sinks of the six IGBT switching transistors inside the power source to collect the surface temperature of the different switching transistors. Impedance angle acquisition module: The load end is connected in parallel with an impedance analyzer, which uploads the load impedance angle data in real time via RS485 bus; The collected instantaneous values ​​of three-phase voltage (Ua, Ub, Uc), effective values ​​of three-phase current (Ia, Ib, Ic), surface temperatures of 6 switching transistors, and load impedance angle are aligned according to timestamps to form a four-dimensional time-series data sequence with dimensions N×13; where N is the number of sampling points, and 13 includes 3 voltages, 3 currents, 6 temperatures, and 1 impedance angle. The four-dimensional time series data sequence is preprocessed, including removing DC components from the voltage and current sequences, performing median filtering on the temperature sequence, and extracting high-frequency transient features by wavelet packet decomposition. The dimension is 16, which are all existing conventional technical solutions. The specific implementation steps are not described here. When fusing spatiotemporal correlations through an attention mechanism, the mutual information entropy I between voltage transient features and temperature drift features is calculated. The expression involved is: ,Y). ;in, Y and Y are both random variables, corresponding to voltage transients and temperature drift; p(x,y) is the joint probability density, representing the random variable. The joint probability of taking x and Y taking y is the probability that the voltage transient characteristic x and the temperature drift characteristic y occur simultaneously. p(x) and p(y) are the marginal probability densities, i.e., when Y is not considered. The probability of taking x is not considered. The probability that Y takes the value y; It is the log-likelihood ratio; When I( When Y)≥0.8, the feature is determined to be strongly correlated, and a weight coefficient w=1.2 is assigned. When 0.3 <I( When (Y) < 0.8, assign a weighting coefficient w = 1.0; When I( When (,Y)≤0.3, assign a weighting coefficient w=0.5 to suppress weak correlation noise; When constructing the dynamic feature matrix, the extracted 16-dimensional high-frequency transient features, 6-dimensional temperature drift features, and 1-dimensional impedance angle features are weighted and summed to form a dynamic feature matrix of dimension N×23, where N is the number of time steps. Where X is the high-frequency transient feature matrix with a dimension of N×16; is the high-frequency transient feature weight vector, obtained by calculating and analyzing mutual information entropy; Y is the temperature drift feature matrix with a dimension of N×6. This is the temperature drift feature weight vector, and... The mutual information entropy is calculated based on temperature characteristics and voltage entropy; Z is the impedance angle characteristic matrix with a dimension of N×1; This is the characteristic weight of the impedance angle, with a default value of 1. This indicates element-wise multiplication, where each feature dimension is multiplied by its corresponding weight. This indicates horizontal concatenation, where the weighted feature matrices are concatenated column-wise to form the total matrix.

[0022] It should be noted that the precise deployment of multimodal sensors provides high signal-to-noise ratio raw data for subsequent feature extraction, and the synchronized time stamp ensures the accuracy of spatiotemporal correlation analysis; The four-layer decomposition specifically covers power switching transients, and the energy entropy quantifies the intensity of high-frequency disturbances, which can solve the problem that traditional Fourier transform cannot capture non-stationary signals. By dynamically weighting the mutual information entropy, the strong correlation between voltage transients and temperature drift is highlighted. For example, the temperature rise caused by IGBT switching losses is coupled with voltage ripple. Weak correlation noise from environmental interference is suppressed, which can improve the discriminativeness of the feature matrix.

[0023] In this embodiment of the invention, by fusing temperature and impedance angle information, compared with traditional single physical quantity acquisition, the risk of thermal runaway of switching transistors can be effectively warned in advance; by using the attention mechanism, the signal-to-noise ratio of the feature matrix is ​​effectively improved, and the output delay of the control strategy of the subsequent AI model is reduced.

[0024] The generated dynamic feature matrix is ​​input into a pre-trained deep reinforcement learning model, which outputs a dual-time-scale control strategy containing short-time voltage compensation and long-term parameter adjustment. The short-time voltage compensation uses an LSTM sub-network to predict future voltage fluctuations and generate a PWM duty cycle correction value. The long-term parameter adjustment is based on the temperature drift trend in the feature matrix, and the proportional gain and integral time of the PID controller are optimized using a gradient descent algorithm. The specific steps include: The pre-trained deep reinforcement learning model adopts the DDPG framework, which includes an Actor network and a Critic network. The Actor network corresponds to the policy network, and the Critic network corresponds to the value network. Both of them adopt a fully connected neural network structure. The Actor network consists of an input layer, hidden layer 1, hidden layer 2, and an output layer. Input layer: dynamic feature matrix with dimensions 23×1; Hidden layer 1: 128 neurons, ReLU activated; Hidden layer 2: 64 neurons, ReLU activated; Output layer: 2 neurons, Tanh activation, output PWM duty cycle correction value PID controller parameter correction amount Integral time correction ; The Critic network consists of an input layer, hidden layer 1, hidden layer 2, and an output layer; Input layer: states and actions, with 23+3=26 dimensions; Hidden layer 1: 256 neurons, ReLU activated; Hidden layer 2: 128 neurons, ReLU activated; Output layer: 1 neuron, linearly activated, outputs action value Q-value; In addition, define the state space S and the action space A; State space S: Dynamic feature matrix, including voltage transient energy entropy, temperature drift weighted feature, and load impedance angle; Action space A includes short-term actions and long-term actions; Short-time action: PWM duty cycle correction value ∈[−5%,+5%]; Long-term action: PID controller parameter correction ∈[−0.2,+0.2], integration time correction ∈[−0.1s,+0.1s]; The training and optimization of the constructed model are based on existing conventional techniques, and the specific implementation steps will not be elaborated here. When generating short-time voltage compensation through the LSTM sub-network, the input data is the high-frequency transient features in the dynamic feature matrix, which is 16-dimensional and corresponds to the voltage entropy sequence; the time step T=50, that is, the input is the voltage entropy data of the past 50ms. The LSTM subnetwork consists of two LSTM layers, each with 64 neurons, and a dropout rate of 0.2. Fully connected layer: 32 neurons, ReLU activated; Output layer: 1 neuron, linearly activated; Output the predicted voltage fluctuation value within the next 5ms ; It should be noted that the LSTM sub-network processes raw state data with time-series dependencies, such as sensor data from continuous multi-step observations, historical action sequences, etc., and outputs an enhanced state representation that incorporates time-series information. The Actor network takes the timing-enhanced state output from the LSTM subnetwork as input and outputs deterministic actions; a continuous action space, such as PWM duty cycle. In a time-series environment, actions depend not only on the current state but also on the cumulative effects of historical states. For example, in power system voltage control, the current PWM duty cycle adjustment needs to refer to the voltage fluctuation trend over the past five cycles. The LSTM sub-network compresses the original time-series voltage data into low-dimensional time-series features, and the Actor network outputs the duty cycle adjustment based on these features. The Critic network takes the temporal augmented state output by the LSTM subnetwork and the action output by the Actor network as input and evaluates the long-term cumulative reward of the state-action pair, i.e., the Q value. The Critic network needs to determine whether the action chosen by the Actor network is optimal in the current time sequence state. The optimality depends on the trend of the historical state. The time sequence state features provided by the LSTM sub-network enable the Critic network to more accurately evaluate the long-term impact of the action. Therefore, the LSTM sub-network, as a temporal feature extraction module, provides the Actor network and Critic network with a state representation containing historical dependencies. The Actor network and Critic network then complete policy output and value evaluation based on this representation. Together, the three achieve reinforcement learning decision-making for the temporal environment. Based on the predicted voltage fluctuation, the PWM duty cycle correction value is calculated and generated. The relevant expression is as follows: ;in, This is a proportionality coefficient, which can be determined based on previous experimental data; the default value is 0.02. is the integration coefficient, with a default value of 0.01; t is the integration time window; During real-time inference, the dynamic feature matrix is ​​truncated into the LSTM subnetwork via a sliding window; the window size is 50 and the stride is 1. Voltage fluctuation prediction value is output every 1ms. And calculate the PWM duty cycle correction value. ; The correction value is injected into the PWM generator of the power source drive module through a field-programmable gate array; Furthermore, when generating long-term parameter adjustments through gradient descent optimization, the loss function is defined with the output voltage drift rate δU over 24 hours as the optimization objective: ;in, , where are the PID parameters to be optimized. The integration time; These are the proportional gain regularization coefficient and the integral time regularization coefficient, respectively, with default values ​​of 0.01 and 0.005. The output voltage drift rate over 24 hours. , These are the rated output voltage, the maximum voltage value within 24 hours, and the minimum voltage value, respectively. The square of the L2 norm; The Adam optimizer is used to update PID parameters, and the relevant expressions are: ;in, The PID parameters are for the j-th iteration; Here are the updated parameters for the (j+1)th iteration; j is the iteration index. For the loss function in gradient at; The learning rate is set to 0.001 by default. Wherein, the initial PID parameters correspond to , ; The frequency is adjusted to perform gradient descent optimization once per hour, based on the dynamic feature matrix of the past hour.

[0025] In this embodiment of the invention, the prediction error of voltage fluctuation can be effectively reduced by using the LSTM sub-network, and ripple can be compensated in advance, thereby reducing the output voltage ripple coefficient; gradient descent optimization can shorten the PID parameter adjustment time and effectively control the 24-hour drift rate; the DDPG model has low output delay of the control strategy during load mutation, and the system recovery stabilization time can be effectively reduced compared with PI control, thus enhancing robustness.

[0026] The output PWM duty cycle correction value is injected into the power source drive module, and the compensated output voltage ripple coefficient and drift rate are collected simultaneously. Data analysis is performed on the voltage ripple coefficient and drift rate to determine whether the high-power source output is balanced. If unbalanced, the parameter optimization process is re-executed until the analysis result indicates that the high-power source output is balanced. Specific steps include: The output PWM duty cycle correction value is injected into the power source drive module through a high-speed digital interface; the high-speed digital interface is, for example, SPI or EtherCAT, with a communication rate ≥1Mbps; the power source drive module is, for example, an IGBT inverter drive circuit. During injection, a hardware synchronization signal is used to ensure that the correction value is strictly synchronized with the power source switching cycle, avoiding voltage fluctuations caused by timing deviations; the hardware synchronization signal is, for example, a 10kHz synchronization pulse generated by the FPGA; Among them, the interface communication delay is ≤10μs to ensure that the correction value takes effect in real time; The synchronous pulse has a 50% duty cycle, is injected on the rising edge, and is aligned with the power transistor's turn-on time. The compensated output voltage signal is acquired synchronously using a high-precision voltage probe and a data acquisition card, including ripple coefficient acquisition and 24-hour drift rate acquisition. When acquiring the ripple coefficient, 10 switching cycles are continuously acquired. For example, at a switching frequency of 20kHz, the acquisition time is 0.5ms, and the original voltage waveform data is stored. During 24-hour drift rate acquisition, one RMS voltage value is collected every 100ms using the AD736 true RMS converter chip. This continuous acquisition for 24 hours generates a voltage time series. ,in, For time steps, N = 24 × 3600 × 10 = 864000 data points; The voltage ripple coefficient is calculated based on the peak value of the AC ripple component in the output voltage and the rated voltage. The expression involved is: ;in, The peak-to-peak value of the voltage waveform over 10 switching cycles; This is the rated output voltage of the power source; And, based on the voltage time series generated over 24 hours of continuous data acquisition. Calculate the drift rate for the corresponding 24 hours. ; The calculated voltage ripple coefficient Drift rate The values ​​of the first and second balance thresholds are compared and judged respectively. The values ​​of the first and second balance thresholds can be obtained through simulation tests based on previous sample data, or can be customized by professionals in this field based on their work experience and requirements. The specific values ​​are not limited. If the voltage ripple coefficient is less than or equal to the first balance threshold and the absolute value of the drift rate is less than or equal to the second balance threshold, then the high power source output is considered balanced. If the voltage ripple coefficient is greater than the first balance threshold and / or the drift rate is greater than the second balance threshold, the high power source output is determined to be unbalanced and an interrupt signal is generated. The parameter optimization process is re-executed through the interrupt signal to generate a new PWM duty cycle correction value, and the compensation and determination are repeated.

[0027] In this embodiment of the invention, by injecting the output PWM duty cycle correction value into the power source drive module through a high-speed digital interface, the timing consistency between the PWM correction value and the voltage sampling can be ensured, effectively reducing the measurement error of the ripple coefficient. By simultaneously monitoring the high-frequency ripple of short-term stability and the long-term drift of long-term stability, the power source output quality can be reflected more comprehensively and reliably than the existing single index. By analyzing the voltage ripple coefficient and drift rate, it is determined whether the high power source output is balanced, and the parameter optimization process is dynamically triggered to re-execute, realizing automated closed-loop parameter tuning, adapting to fast dynamic load scenarios, and further improving the reliability and adaptability of the balance.

[0028] In the several embodiments provided by this invention, it should be understood that the disclosed system can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative; for example, the division of modules is only a logical functional division, and there may be other division methods in actual implementation.

[0029] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0030] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in the form of hardware plus software functional modules.

[0031] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the essential characteristics of the present invention.

[0032] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A power source balancing method based on AI, characterized in that, include: A multi-mode measurement component is deployed between the output end and the load end of a high-precision three-phase power source to synchronously acquire the instantaneous values ​​of three-phase voltage, the effective values ​​of current, the junction temperature of the switching transistor, and the load impedance angle, generating a four-dimensional time-series data sequence. The four-dimensional time-series data sequence is preprocessed, and the spatiotemporal correlation is fused through an attention mechanism to construct a dynamic feature matrix. The generated dynamic feature matrix is ​​input into a pre-trained deep reinforcement learning model, and the output is a dual-time-scaled control strategy containing short-time voltage compensation and long-term parameter adjustment. The short-time voltage compensation is generated by predicting future voltage fluctuations through an LSTM sub-network to generate a PWM duty cycle correction value. The long-term parameter adjustment is based on the temperature drift trend in the feature matrix and optimizes the proportional coefficient and integral time of the PID controller through a gradient descent algorithm. The output PWM duty cycle correction value is injected into the power source drive module, and the compensated output voltage ripple coefficient and drift rate are collected synchronously. The voltage ripple coefficient and drift rate are analyzed to determine whether the high power source output is unbalanced. If unbalanced, the parameter optimization process is re-executed until the analysis result shows that the high power source output is balanced.

2. The AI-based power source balancing method according to claim 1, characterized in that, A measurement component consisting of a voltage acquisition module, a temperature acquisition module, and an impedance angle acquisition module is deployed in series between the output terminal and the load terminal of a high-precision three-phase power source. The collected instantaneous values ​​of three-phase voltage, effective values ​​of three-phase current, surface temperatures of six switching transistors, and load impedance angles are aligned according to timestamps to form a four-dimensional time-series data sequence.

3. The AI-based power source balancing method according to claim 2, characterized in that, When the spatiotemporal correlation is fused through the attention mechanism, the mutual information entropy of voltage transient characteristics and temperature drift characteristics is calculated; data analysis is performed on the mutual information entropy, and weight coefficients are dynamically assigned. When constructing the dynamic feature matrix, the extracted 16-dimensional high-frequency transient features, 6-dimensional temperature drift features, and 1-dimensional impedance angle features are weighted and summed to form the dynamic feature matrix.

4. The AI-based power source balancing method according to claim 3, characterized in that, The pre-trained deep reinforcement learning model adopts the DDPG framework, which includes an Actor network and a Critic network. The Actor network corresponds to the policy network, and the Critic network corresponds to the value network. Both adopt a fully connected neural network structure.

5. The AI-based power source balancing method according to claim 4, characterized in that, Based on the predicted voltage fluctuations The PWM duty cycle correction value is generated by calculation, and the relevant expression is as follows: ;in, This is the proportionality coefficient; is the integration coefficient, with a default value of 0.01; t is the integration time window.

6. The AI-based power source balancing method according to claim 5, characterized in that, When optimizing the long-term parameter adjustment using gradient descent, the output voltage drift rate δU over 24 hours is used as the optimization objective, and the loss function is defined as follows: ;in, , where are the PID parameters to be optimized. This is the proportionality coefficient. The integration time; These are the proportional coefficient regularization coefficient and the integral time regularization coefficient, respectively. The output voltage drift rate over 24 hours. , These are the rated output voltage, the maximum voltage value within 24 hours, and the minimum voltage value, respectively. It is the square of the L2 norm.

7. The AI-based power source balancing method according to claim 6, characterized in that, Updating PID parameters using the Adam optimizer involves the following expressions: ;in, The PID parameters are for the j-th iteration; Here are the updated parameters for the (j+1)th iteration; j is the iteration index. For the loss function in gradient at; This is the learning rate.

8. The AI-based power source balancing method according to claim 7, characterized in that, The output PWM duty cycle correction value is injected into the power source drive module through a high-speed digital interface; during injection, a hardware synchronization signal is used to ensure that the correction value is strictly synchronized with the power source switching cycle to avoid voltage fluctuations caused by timing deviations.

9. The AI-based power source balancing method according to claim 8, characterized in that, When collecting drift rate data for 24 hours, one effective voltage value is collected every 100ms, and the data is continuously collected for 24 hours to generate a voltage time series. The voltage ripple coefficient is calculated based on the peak value of the AC ripple component in the output voltage and the rated voltage.

10. A power source balancing method based on AI according to claim 9, characterized in that, The calculated voltage ripple coefficient and drift rate are compared with the corresponding first balance threshold and second balance threshold, respectively. If the voltage ripple coefficient is less than or equal to the first balance threshold and the absolute value of the drift rate is less than or equal to the second balance threshold, then the high power source output is considered balanced. If the voltage ripple coefficient is greater than the first balance threshold and / or the drift rate is greater than the second balance threshold, the high power source output is determined to be unbalanced and an interrupt signal is generated. The parameter optimization process is re-executed through the interrupt signal to generate a new PWM duty cycle correction value, and the compensation and determination are repeated.

Images