Intelligent capacitor bank balance optimization control system and method

Abstract

The invention relates to the technical field of power electronic intelligent control, and discloses an intelligent capacitor bank balance optimization control system and method.The intelligent capacitor bank balance optimization control method comprises the steps that real-time parameters of all units of a capacitor bank are collected; constructing a multi-dimensional imbalance evaluation model, and quantifying the imbalance state of the capacitor bank; based on the quantification result, combining deep reinforcement learning and double-layer closed-loop control to generate an adaptive balance control strategy; based on a self-adaptive balance control strategy, realizing accurate control on the capacitor bank through a power electronic interface, and establishing a closed-loop feedback mechanism to obtain a control execution result and a new system state; based on the new state and historical operation data of the system, the health state of the capacitor bank is evaluated, the remaining service life is predicted, and a predictive maintenance strategy is formulated, the problem of voltage imbalance of the capacitor bank in long-term operation can be solved, and efficient utilization and service life prolonging of the capacitor bank are achieved.

Description

Technical Field

[0001] This invention relates to the field of power electronics intelligent control technology, and more specifically, to an intelligent capacitor bank balance optimization control system and method. Background Technology

[0002] In the field of electronic intelligent control, capacitor banks, as crucial components for energy storage and filtering, directly impact the stability and efficiency of the entire system. However, in practical applications, especially when capacitor banks are used in parallel, voltage imbalances often occur due to manufacturing differences, variations in environmental conditions, and long-term operation. This not only leads to premature failure of some units but also causes a decline in overall system performance. Furthermore, with prolonged operation, the inconsistent rate of capacitance decay among different units exacerbates this imbalance, further affecting charging and discharging efficiency and lifespan.

[0003] To address these issues, traditional control methods often rely on simple hardware balancing circuits or software algorithms based on empirical rules. While these methods can alleviate imbalances to some extent, they struggle to accurately assess and optimize the capacitor bank's condition in the face of complex and ever-changing real-world operating conditions. Especially when considering capacitor bank health monitoring and predictive maintenance, existing technologies fall short. For example, traditional methods typically cannot accurately predict remaining service life, leading to inappropriate maintenance timing and increasing unnecessary costs or the occurrence of risky failures.

[0004] To address these issues, a more intelligent and adaptive control system is needed to manage the operating status of the capacitor bank. The ideal solution should be able to monitor the key parameters of each capacitor cell in real time and construct a multi-dimensional imbalance assessment model accordingly, thereby achieving accurate quantitative analysis of the imbalance state. Furthermore, the system should incorporate advanced machine learning algorithms, such as deep reinforcement learning, to generate adaptive balance control strategies, ensuring efficient operation of the capacitor bank under various operating conditions while extending its service life. Summary of the Invention

[0005] This invention provides an intelligent capacitor bank balance optimization control system and method, which solves the technical problems of voltage imbalance, inconsistent capacity decay, low charging and discharging efficiency, and inaccurate life prediction in the parallel use of capacitor banks in related technologies.

[0006] This invention provides an intelligent capacitor bank balance optimization control method, comprising the following steps:

[0007] Real-time parameters of each unit in the capacitor bank are collected through a distributed sensor network to establish a dynamic data model of the capacitor bank.

[0008] Based on the real-time parameters of each unit of the capacitor bank collected, a multi-dimensional imbalance assessment model is constructed to quantify the imbalance state of the capacitor bank.

[0009] Based on the quantification results of the unbalanced state of the capacitor bank, an adaptive balance control strategy is generated by combining deep reinforcement learning and two-layer closed-loop control.

[0010] Based on the generated adaptive balance control strategy, the capacitor bank is precisely controlled through the power electronic interface, and a closed-loop feedback mechanism is established to obtain the control execution results and the new system state.

[0011] Based on the new system status and historical operating data, assess the health status of the capacitor bank, predict its remaining service life, and develop predictive maintenance strategies.

[0012] In a preferred embodiment, real-time parameters of each unit of the capacitor bank are collected through a distributed sensor network to establish a dynamic data model of the capacitor bank, including:

[0013] For each capacitor cell C in the capacitor bank i (i=1,2,...,n), high-precision sensors are used to collect voltage, current and surface temperature and internal resistance values, where n represents the number of capacitor units;

[0014] An adaptive sampling frequency algorithm based on capacitance characteristics is adopted to automatically adjust the sampling frequency according to the state of the capacitor bank. The sampling frequency is a weighted combination of the base sampling frequency and the standard deviations of voltage, current and temperature.

[0015] Based on the collected basic parameters, characteristic parameters are calculated, including effective capacity, self-discharge rate, temperature gradient, and internal resistance change rate.

[0016] A multi-parameter dynamic correlation model is constructed, and the early identification of multi-parameter abnormal states is achieved by calculating the dynamic correlation matrix between parameters.

[0017] In a preferred embodiment, a multi-dimensional imbalance assessment model is constructed based on the real-time parameters of each unit of the capacitor bank to quantify the imbalance state of the capacitor bank, including:

[0018] Calculate single-parameter imbalance, including voltage imbalance, temperature imbalance, capacity imbalance, and internal resistance imbalance;

[0019] The parameter weights are adaptively determined based on the entropy weight method by calculating the information entropy, difference coefficient and weight of the parameters.

[0020] The overall imbalance degree is calculated based on the adaptive weight adjustment of the working conditions. The overall imbalance degree is the weighted sum of the imbalance degrees of each individual parameter, and the weights are dynamically adjusted according to the working conditions.

[0021] An improved fuzzy C-means clustering algorithm is used to classify the imbalance state into voltage imbalance, temperature imbalance, capacity imbalance, internal resistance imbalance and mixed type.

[0022] The trend of imbalance is calculated using the adaptive exponential weighted moving average method, and the future trend of imbalance is predicted using the ARIMA model.

[0023] In a preferred embodiment, based on the quantization results of the capacitor bank's unbalanced state, an adaptive balance control strategy is generated by combining deep reinforcement learning and two-layer closed-loop control, including:

[0024] A two-layer control architecture is constructed. The inner layer control is based on an enhanced PID controller to achieve millisecond-level fast response. The PID controller includes proportional, integral, derivative and nonlinear correction terms.

[0025] The outer control layer achieves long-term optimized scheduling based on a fusion of deep reinforcement learning and MPC control;

[0026] An improved deep deterministic strategy gradient algorithm is adopted, including a multi-scale time-aware Actor network structure and a two-stream Critic network structure;

[0027] Design a dynamic adaptive fusion mechanism for deep reinforcement learning and model predictive control, and dynamically adjust the weights of the two control strategies through adaptive fusion coefficients;

[0028] A recursive least squares method combining particle filtering is used to achieve high-precision identification of time-varying system parameters.

[0029] In a preferred embodiment, the improved deep deterministic policy gradient algorithm includes:

[0030] The state space is defined as a vector containing voltage, current, temperature, internal resistance, state of charge, and unbalance parameters.

[0031] The action space is defined as a vector containing charging current, discharging current, charging time, discharging time, and control mode;

[0032] The multi-objective reward function is designed as a weighted combination to reduce overall imbalance, energy loss and maximum temperature rise, while improving energy efficiency and reducing health status loss.

[0033] An improved priority experience replay mechanism is adopted, and the priority calculation takes into account the impact of TD error and the number of times the sample is sampled.

[0034] In a preferred embodiment, based on the generated adaptive balance control strategy, precise control of the capacitor bank is achieved through a power electronics interface, including:

[0035] The control strategy is converted into a PWM control signal, and an adaptive duty cycle mapping algorithm is adopted. The duty cycle is related to the current value, voltage and temperature.

[0036] S-curve trajectory planning is used to generate smooth control signals, avoiding electrical shocks caused by sudden changes in control signals;

[0037] Multi-layer closed-loop control is achieved. The inner loop control adopts an active disturbance rejection current control algorithm with a control period of 5μs, while the outer loop control adopts model-based predictive control with a control period of 500μs.

[0038] It adopts a bidirectional DC-DC converter topology, supports 0-5A continuously adjustable charging and discharging current, and combines soft-switching technology with zero voltage turn-on and zero current turn-off.

[0039] Temperature prediction control based on neural networks enables intelligent thermal management, including an active thermal runaway prevention system with gradient response.

[0040] In a preferred embodiment, the health status of the capacitor bank is assessed and its remaining service life is predicted based on the new state of the system and historical operating data, including:

[0041] Calculate multi-dimensional health status indicators, including health status indices based on capacity, internal resistance, self-discharge rate, and temperature response characteristics;

[0042] A multi-dimensional health index is integrated using dynamic weighting coefficients, which are adaptively adjusted based on current working conditions and historical data.

[0043] The remaining useful life is predicted based on a hybrid deep learning architecture of LSTM and residual network. The input feature set includes static features, temporal features, derived features and external factors.

[0044] A comprehensive loss function combining mean square error, quantile loss, and uncertainty loss is adopted;

[0045] The final predicted output is the output value of the residual network model and its uncertainty.

[0046] In a preferred embodiment, a smart capacitor bank balance optimization control method further includes:

[0047] A multi-source failure mode library is constructed, including failure modes such as capacity decay, internal resistance increase, self-discharge increase, temperature anomaly, and balance performance deterioration.

[0048] A multi-class diagnostic system employing an ensemble learning framework integrates classification results from cascaded random forests, support vector machines, gradient boosting decision trees, and fully connected neural networks.

[0049] An adaptive anomaly detection algorithm combining local anomaly factors and isolated forests is used to calculate anomaly scores by weighted fusion of the results of the two algorithms.

[0050] A three-level gradient early warning mechanism is generated based on the fault risk level;

[0051] Based on the predictive lifetime-aware balance control strategy, the control input is calculated as a function product of the unit's baseline control quantity, health status, and remaining lifetime, ensuring that capacitor units with poor health status or short remaining lifetime are more strictly protected.

[0052] In a preferred embodiment, developing a predictive maintenance strategy includes:

[0053] The optimal maintenance timing is determined based on a multi-objective lifetime-cost optimization model, which is achieved by minimizing the weighted sum of maintenance cost, failure risk cost, and operational impact cost.

[0054] Calculate maintenance priority; the risk score is the product of failure probability, failure severity, and maintenance urgency.

[0055] Based on clustering algorithms, cells with similar remaining lifespans are grouped together, and a group optimization maintenance strategy is implemented.

[0056] Based on the unit's health status, fault type, and system operation requirements, replacement suggestions, balancing adjustment suggestions, parameter recalibration suggestions, deep balancing suggestions, and preventive detection suggestions are generated. After maintenance is performed, execution effect data is collected, and the health status assessment and prediction models are updated.

[0057] In a preferred embodiment, an intelligent capacitor bank balancing optimization control system is used to execute the steps of the intelligent capacitor bank balancing optimization control method described above, including:

[0058] A distributed sensor network is used to collect the voltage, current, surface temperature, and internal resistance of each unit in the capacitor bank.

[0059] The edge layer computing unit is used for data processing, feature parameter calculation and imbalance evaluation. It uses the time series database InfluxDB to store data and sets up a hierarchical caching mechanism.

[0060] The cloud-based intelligent decision-making center includes a deep reinforcement learning module and a model prediction control module, used to generate adaptive equilibrium control strategies;

[0061] The power electronic interface adopts a bidirectional DC-DC converter topology and integrates SiCMOSFET drive technology and soft-switching technology.

[0062] A closed-loop feedback controller enables multi-layer closed-loop control, including inner-loop active disturbance rejection current control and outer-loop predictive control.

[0063] The health status assessment and prediction module, based on a hybrid deep learning architecture of LSTM and residual network, realizes health status assessment and remaining useful life prediction.

[0064] The predictive maintenance decision support module is used for fault mode identification, anomaly detection, early warning generation, and maintenance strategy formulation.

[0065] The beneficial effects of this invention are as follows:

[0066] Key parameters of each unit in the capacitor bank are collected in real time through a distributed sensor network, and the imbalance state is quantified using a multi-dimensional imbalance evaluation model. Based on this, an adaptive balance control strategy is generated by combining deep reinforcement learning and a two-layer closed-loop control mechanism, achieving precise regulation of the capacitor bank.

[0067] Furthermore, it can assess the health status of the capacitor bank and predict its remaining service life based on the system's new status and historical operating data, thus developing predictive maintenance strategies. This not only effectively solves the voltage imbalance problem in the long-term operation of the capacitor bank but also improves the system's reliability and efficiency while reducing maintenance costs. Attached Figure Description

[0068] Figure 1 This is a flowchart of an intelligent capacitor bank balance optimization control method according to the present invention. Detailed Implementation

[0069] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0070] At least one embodiment of the present invention discloses an intelligent capacitor bank balance optimization control method, such as... Figure 1 As shown, it includes the following steps:

[0071] S100 collects real-time parameters of each unit in the capacitor bank through a distributed sensor network and establishes a dynamic data model of the capacitor bank.

[0072] This step employs a high anti-interference design and sensor data fusion technology to ensure the accuracy and stability of parameter acquisition, and specifically includes the following steps:

[0073] S110, parameter acquisition;

[0074] For each capacitor cell C in the capacitor banki (i = 1, 2, ..., n), voltage values ​​V are collected using high-precision sensors. i (t), Current value I i (t), surface temperature T i (t) and internal resistance R i (t), where n represents the total number of capacitor units. The sensor arrangement adopts a symmetrical structure design, with 4 temperature sensing points set in each capacitor unit to ensure the representativeness of the parameter acquisition.

[0075] This embodiment employs an adaptive sampling frequency algorithm based on capacitance characteristics, which automatically adjusts the sampling frequency according to the state of the capacitor bank:

[0076] f sample (t)=f base ·(1+k v ·σ V (t)+k i ·σ I (t)+k t ·σ T (t))

[0077] Where f base Based on the sampling frequency; σ V (t), σ I (t), σ T (t) represents the standard deviation of voltage, current, and temperature, respectively; k v k i k t These are the first, second, and third weighting coefficients, used to adjust the degree of influence of each parameter on the sampling frequency; f sample (t) represents the adaptive sampling frequency.

[0078] S120, Data Processing;

[0079] A sliding window averaging method is used to reduce the impact of noise. Data smoothing is achieved by calculating the average of the sampled values ​​at the current time point and the previous M-1 time points. Here, M is the size of the sliding window.

[0080] Outlier detection uses the 3σ criterion: when the deviation of a parameter value from the mean exceeds three times the standard deviation, it is judged as an outlier and is removed or corrected.

[0081] Data standardization is performed using the min-max normalization method, which maps parameter values ​​to the [0, 1] interval to facilitate subsequent processing.

[0082] S130, Calculation of characteristic parameters;

[0083] Based on the collected basic parameters, the following characteristic parameters are calculated: Effective capacity: Calculated by the relationship between the current integral and voltage change of the capacitor unit during the charging and discharging process, reflecting the actual energy storage capacity of the capacitor unit;

[0084] Self-discharge rate: determined by measuring the natural rate of decrease of the capacitor cell voltage under no external current, reflecting the energy retention capability of the capacitor cell;

[0085] Temperature gradient: Calculates the rate of change of surface temperature of capacitor cell per unit time, reflecting changes in thermal properties;

[0086] Internal resistance change rate: The relative change of the current internal resistance value with the resistance value within the reference time, reflecting the change of the internal state of the capacitor cell.

[0087] S140, multi-parameter dynamic correlation;

[0088] This embodiment designs a multi-parameter abnormal state early identification algorithm, which calculates the dynamic correlation matrix between parameters:

[0089]

[0090] Where R corr (t) represents the dynamic correlation matrix between parameters; where ρ XY (t) represents the dynamic correlation coefficient of parameters X and Y within the time window t.

[0091] S150, data transmission and storage;

[0092] The system uses CAN bus communication at a transmission rate of 500kbps to transmit data to the edge layer computing unit 40. Data storage utilizes the time-series database Influx DB, with a hierarchical caching mechanism to prevent data loss due to communication interruptions. AES-256 encryption is employed during transmission to ensure data security.

[0093] S200, based on the real-time parameters of each unit of the capacitor bank, constructs a multi-dimensional imbalance assessment model to quantify the imbalance state of the capacitor bank;

[0094] Based on the parameters obtained from S100, a multi-dimensional imbalance assessment model is constructed to quantify the imbalance state of the capacitor bank. Through adaptive weight allocation and dynamic threshold adjustment, accurate assessment of the imbalance state under different operating conditions is achieved; specifically, the following steps are included:

[0095] S210, Single-parameter unbalance calculation;

[0096] Voltage imbalance is expressed using the coefficient of variation, which represents the degree of dispersion in the voltage distribution.

[0097]

[0098] in This represents the average voltage across all capacitor cells; V i U represents the voltage value of the i-th capacitor cell; n represents the total number of capacitor cells; U V This indicates the degree of voltage imbalance.

[0099] Temperature imbalance:

[0100]

[0101] Where T max Indicates the highest temperature; T min Indicates the lowest temperature; Indicates average temperature; T i This represents the temperature of the i-th capacitor cell; β T This represents the temperature distribution shape factor, used to adjust the effect of the temperature distribution shape on the degree of imbalance; the default value is 0.5. T This represents the temperature imbalance; n represents the total number of capacitor cells. This embodiment innovatively introduces a distribution shape factor to achieve precise quantification of various temperature distribution patterns.

[0102] Capacity imbalance:

[0103]

[0104] Where C max Indicates maximum capacity; C min Indicates minimum capacity; C rated Indicates rated capacity; C i This represents the capacitance of the i-th capacitor unit; Indicates average capacity; U C This indicates the degree of capacitance imbalance. This index takes into account both extreme value differences and the degree of distribution dispersion; n represents the total number of capacitor cells.

[0105] Internal resistance imbalance, a weighted combination of the coefficient of variation of the internal resistance distribution and the relative growth rate:

[0106]

[0107] in R represents the average internal resistance; i γ represents the internal resistance of the i-th capacitor unit; R This represents the internal resistance weighting coefficient, with a default value of 0.3; ΔR i U represents the rate of change of internal resistance; R This indicates the degree of internal resistance imbalance; n represents the total number of capacitor cells. This index combines static imbalance with dynamic trends for evaluation.

[0108] S220, parameter weights are adaptively determined based on the entropy weight method;

[0109] Data standardization processing:

[0110]

[0111] Where x ij r represents the original value of the i-th sample on the j-th indicator; ij This represents the standardized value; This represents the minimum value of all samples for the i-th indicator; This represents the maximum value of all samples for the i-th indicator;

[0112] Calculate the information entropy of the j-th parameter:

[0113]

[0114] in e represents the weight of the i-th sample on the j-th indicator; j The information entropy of the j-th parameter is represented by n; n represents the total number of capacitor units.

[0115] Calculate the coefficient of difference:

[0116] d j =1-e j

[0117] Where d j e represents the difference coefficient of the j-th parameter, reflecting the degree of information provided by that parameter; j Let represent the information entropy of the j-th parameter.

[0118] Calculate the weights:

[0119]

[0120] Where m represents the number of parameters, in this embodiment m = 4, corresponding to four parameters: voltage, temperature, capacity, and internal resistance; d j w represents the difference coefficient of the j-th parameter; j This represents the weight of the j-th parameter.

[0121] S230, Dynamic operating condition adaptive comprehensive unbalance calculation;

[0122] Overall imbalance based on adaptive weight adjustment for operating conditions:

[0123]

[0124] Among them U j The j-th single-parameter imbalance is represented by m; the number of parameters is represented by w. j w represents the weight of the j-th parameter;j (C) indicates the adaptive weights considering operating condition C; U total This indicates the overall degree of imbalance.

[0125] Working condition adaptive weight calculation function:

[0126]

[0127] Where w j w represents the weight of the j-th parameter; j (C) indicates the adaptive weights considering operating condition C; f k (C,j) represents the influence function of the k-th working condition on the weight of the j-th imbalance; δ k This represents the strength coefficient of the working condition influence; K represents the number of working condition types considered.

[0128] Example of influence function for typical operating conditions:

[0129] High-temperature environment conditions (k=1):

[0130]

[0131] Where f1(C,j) represents the effect function of high-temperature environmental conditions on temperature; T amb Indicates ambient temperature.

[0132] High current operating conditions (k=2):

[0133]

[0134] Where f2(C,j) represents the effect function of high current condition on voltage; I represents the current value; I rated Indicates the rated current.

[0135] High-frequency charge and discharge conditions (k=3):

[0136]

[0137] Where f3(C,j) represents the effect function of high-frequency charging and discharging conditions on capacity.

[0138] S240, intelligent identification of imbalance types;

[0139] An improved fuzzy C-means clustering algorithm is used to classify imbalance states into voltage imbalance, temperature imbalance, capacity imbalance, internal resistance imbalance, and mixed imbalance. Its innovation lies in the introduction of an adaptive distance metric and a semi-supervised learning mechanism.

[0140] Improved distance metric:

[0141]

[0142] Where x ik Represents the value of sample i in the k-th feature dimension; m represents the number of parameters; v jk λ represents the value of the j-th cluster center on the k-th feature; k (U k ) represents the adaptive feature weights based on imbalance, calculated using the following formula: d ij U represents the weighted distance from sample i to cluster center j; k U represents the imbalance of the k-th parameter; l This represents the degree of imbalance of the l-th parameter.

[0143] Membership degree calculation formula:

[0144]

[0145] Where m fuzzy represents the fuzziness factor, controlling the degree of fuzziness in the clustering results; in this embodiment, it is set to 2; c represents the number of cluster centers; in this embodiment, c = 5, corresponding to the five imbalance types; μ ij d represents the membership degree of sample i to cluster j, with a value range of [0, 1]; ik d represents the distance from sample i to cluster center k; ij This represents the distance from sample i to cluster center j.

[0146] Clustering center update:

[0147]

[0148] Where v j x represents the j-th cluster center; i This represents the i-th sample point; n represents the total number of capacitor units; m fuzzy μ represents the fuzziness factor. ij This represents the membership degree of sample i to cluster j.

[0149] Termination conditions:

[0150]

[0151] in represents the membership degree of sample i to cluster j at the t-th and t-1-th iterations, respectively; ∈ represents the convergence threshold, which is 0.001 in this embodiment; This represents calculating the maximum value of the change in membership degree among all combinations of sample i and cluster j.

[0152] S250, Imbalance Trend Analysis and Forecast;

[0153] An adaptive exponential weighted moving average method is used to calculate the trend of imbalance. The smoothing factor is adaptively adjusted according to the difference between the current imbalance and the trend value, so as to accurately track different rates of change.

[0154] Trend prediction: The ARIMA model is used to predict future imbalance trends. The autocorrelation, trend and random fluctuation characteristics of the data are captured by three parts: autoregression, difference and moving average.

[0155] Trend Judgment and Early Warning: By comparing the historical trend change with the predicted trend change, when both are positive, it indicates that the imbalance is continuing to worsen, and the system issues a trend warning.

[0156] S300, based on the unbalanced state of the quantized capacitor bank, combines deep reinforcement learning and two-layer closed-loop control to generate an adaptive balance control strategy;

[0157] This embodiment proposes a fusion mechanism of reinforcement learning and model predictive control, which enables intelligent and precise control of capacitor banks.

[0158] S310 features a dual-layer control architecture design.

[0159] Inner layer control, based on an enhanced PID controller, achieves millisecond-level fast response:

[0160]

[0161] Where u PID (t) represents the output value of the PID controller; e(t) represents the control deviation, i.e., the difference between the target value and the actual value; K p This represents the proportional gain coefficient, used to adjust the control signal strength, with a value range of [0, 50, 2.0]; K i K represents the integral gain coefficient, used to eliminate static error, and its value ranges from [0.05, 0.5]; d K represents the differential gain coefficient, used to improve the system response speed, and its value ranges from [0.01, 0.2]; f The nonlinear correction coefficient is represented by f(e(t)); the nonlinear correction function is represented by f(e(t)) in the form of f(e(t)). β1 is the nonlinear exponent; τ represents the integral variable.

[0162] The outer layer control, based on the fusion of deep reinforcement learning and MPC control, achieves long-term optimized scheduling;

[0163] State space S is defined as follows:

[0164] S={V i ,I i ,T i ,Ri SOC i U V U T U C U R U total ,ΔU trend U pred}

[0165] Where V i I i T i R i Let SOC represent the voltage, current, temperature, and internal resistance of the i-th capacitor unit, respectively. i U represents the state of charge of the i-th capacitor cell; V U T U C U R U total These represent voltage, temperature, capacitance, internal resistance, and overall imbalance, respectively; ΔU trend U represents the trend change in the degree of imbalance; pred V represents the predicted value of imbalance. max and V min These represent the maximum and minimum operating voltages of the capacitor unit, respectively.

[0166] Action space A is defined as follows:

[0167] A={I ch,i ,I dis,i ,t ch,i ,t dis,i Mode i}

[0168] Where I ch,i I represents the charging current of the i-th capacitor cell; dis,i Indicates the discharge current; t ch,i Indicates charging time; t dis,i Indicates discharge time; Mode i Indicates the control mode (1-4 represent normal mode, high performance mode, energy saving mode and life extension mode respectively).

[0169] Multi-objective reward function design:

[0170] R(s,a)=-w4·U total -w5·ΔE loss -w6·ΔT max +w7·η-w5·ΔSOH

[0171] Among them U total Indicates the overall imbalance; ΔE lossIndicates energy loss; ΔT max η represents the maximum temperature rise; η represents the energy efficiency; ΔSOH represents the health loss; weight coefficients w4, w5, w6, and w7 represent the fourth, fifth, sixth, and seventh weight coefficients, respectively; s represents the state; a represents the action; and R(s,a) represents the reward function.

[0172] S320 features an innovative dual-layer controller collaborative working mechanism.

[0173] Enhanced MPC control algorithm:

[0174]

[0175] Where u MPC (t) represents the output control quantity of the MPC controller at time t; N p N represents the prediction time domain; c U represents the control time domain; Q represents the output error weight matrix, which is a diagonal matrix; R represents the control increment weight matrix, which is a diagonal matrix; ρ represents the terminal state weight coefficient; y(t+j|t) represents the output predicted from time t to time t+j; r(t+j) represents the reference trajectory; Δu(t+j|t) represents the control increment; U total (t+N p |t) represents the overall imbalance at the end of the prediction time domain; Δu represents the control increment sequence; j represents the time step index within the prediction time domain.

[0176] The inner and outer layer control coordination mechanism uses adaptive weights to achieve smooth fusion based on the control objective and system state.

[0177] u(t)=u PID (t)+γ(t·Δu MPC (t)

[0178]

[0179] Where u(t) represents the final control output; u PID (t) represents the output of the inner PID controller; Δu MPC (t) represents the correction amount of the MPC controller; γ(t) represents the adaptive coordination coefficient; κ represents the switching rate parameter, with a value of 10; U total Indicates the overall imbalance; U threshold This represents the imbalance threshold, which defaults to 0.05; e represents the base of the natural logarithm. This mechanism achieves coordinated control, where the inner layer controls the primary control when the imbalance is low, and the outer layer controls the primary control when the imbalance is high.

[0180] S330, an implementation of the Modified Deep Deterministic Policy Gradient (DDPG) algorithm;

[0181] This embodiment proposes an improved DDPG algorithm, which, considering the characteristics of capacitor bank balance control, introduces time-series correlation learning and multi-scale sensing mechanisms, and has the following features:

[0182] Multi-scale time-aware Actor network structure:

[0183] Input layer: state space dimension;

[0184] Multi-scale feature extraction layer:

[0185] Short timescale branch: 1D convolution, kernel width = 3, output dimension = 128, ReLU activation;

[0186] Mid-timescale branch: 1D convolution, kernel width = 5, output dimension = 128, ReLU activation;

[0187] Long-scale branch: 1D convolution, kernel width = 7, output dimension = 128, ReLU activation;

[0188] Feature fusion layer: fuses features from the three branches through an attention mechanism;

[0189] Hidden layer 1: 256 neurons, ReLU activated;

[0190] Hidden layer 2: 128 neurons, ReLU activated;

[0191] Output layer: Action space dimension, tanh activation;

[0192] Dual-stream Critic network structure:

[0193] State-coded stream:

[0194] Input: Dimension of the state space;

[0195] Hidden layer 1: 256 neurons, ReLU activated;

[0196] Hidden layer 2: 128 neurons, ReLU activated;

[0197] Output: 128-dimensional state encoding vector;

[0198] Action coding stream:

[0199] Input: Action space dimension;

[0200] Hidden layer: 128 neurons, ReLU activated;

[0201] Output: 128-dimensional action encoding vector;

[0202] Fusion layer: connects state coding and action coding;

[0203] Hidden layer: 256 neurons, ReLU activated;

[0204] Improved priority experience replay mechanism:

[0205]

[0206] Where p i Indicates the priority of sample i; δ i Indicates TD error; ∈1 indicates a small positive number to prevent division by zero, and takes a value of 0.01; This represents the priority factor, with a value of 0.6; N i ω represents the number of times sample i is sampled; ω represents the basic sampling bias, with a value of 1; β2 represents the duplication penalty factor, with a value of 0-40, to prevent over-focusing on certain samples.

[0207] Decreasing exploration strategy: An exponentially decaying exploration rate is adopted, with a high initial value to promote exploration, and gradually reduced to a lower stable value as training progresses, so as to achieve a smooth transition from exploration to exploitation.

[0208] Enhanced target network soft update mechanism: The target network parameters are updated by weighted averaging of the original network parameters. The update rate is dynamically adjusted according to the TD error of the current batch. When the error is large, the update is faster, thus improving the convergence speed.

[0209] S340, a deep integration mechanism of DRL and MPC;

[0210] This embodiment proposes a dynamic adaptive fusion mechanism for deep reinforcement learning and model predictive control, achieving complementary advantages between data-driven and model-driven approaches:

[0211] u final (t)=α1(t)·u DRL (t)+(1-α1(t))·u MPC (t)

[0212] Among them, u final (t) represents the final fused control output; u DRL (t) represents the control quantity u output by the deep reinforcement learning algorithm. MPC (t) represents the control quantity output by the MPC algorithm, and α1(t) represents the adaptive fusion coefficient, which takes values ​​in the range [0, 1]. The calculation formula is as follows:

[0213] α1(t)=α base (t)·α confidence (t)·α trend (t)

[0214] Where α base(t) represents the basic fusion coefficient; α confidence (t) represents the confidence level adjustment coefficient; α trend (t) represents the trend adjustment coefficient.

[0215] The overall performance index is calculated by weighting control error, stability, response speed, and health impact.

[0216] S350, adaptive parameter adjustment mechanism;

[0217] Innovative reinforcement learning system parameter identification: A recursive least squares method combined with particle filtering is used to achieve high-precision identification of time-varying system parameters.

[0218] This method updates system parameter estimates through Kalman gain recursion and combines particle filtering to generate correction values, thereby improving parameter tracking capability.

[0219] The adaptive forgetting factor is dynamically adjusted based on the prediction error, enabling the algorithm to better adapt to dynamic changes in the system.

[0220] An adaptive PID parameter tuning algorithm, based on system identification results, employs an improved Ziegler-Nichols rule and combines fuzzy logic to dynamically adjust PID parameters.

[0221]

[0222] K i (t)=f i (K p (t),T u (t),U total (t))

[0223]

[0224] Where K p (t) represents the proportional gain coefficient; K u (t) represents the critical gain under the current operating condition; T u (t) represents the critical period under the current operating condition; U total (t) represents the current overall imbalance; f represents the rate of change of the degree of imbalance; p f i f d K represents the first, second, and third parameter mapping functions based on fuzzy rules, respectively; i (t) represents the integral gain coefficient; K d (t) represents the differential gain coefficient; t represents the current time;

[0225] S360, adaptive selection of control parameters and operating modes;

[0226] The multi-objective mode selection algorithm dynamically selects the optimal operating mode based on system state and external requirements.

[0227]

[0228] Where Mode(t) represents the selected operating mode; μ i (t) represents the weight of the i-th objective, which is dynamically adjusted according to external demand; Score i,j (t) represents the rating of pattern j for target i; N obj The value represents the target quantity; j represents the pattern index; i represents the target index; and t represents the current time.

[0229] The system operating modes include:

[0230] Normal mode: Balances system performance and efficiency, suitable for normal operating environments;

[0231] High-performance mode: Prioritizes system response and stability, suitable for scenarios with high performance requirements;

[0232] Energy-saving mode: Prioritizes energy efficiency and is suitable for energy-constrained scenarios;

[0233] Life Extension Mode: Prioritizes equipment lifespan and is suitable for extending maintenance cycles;

[0234] Emergency mode: A safe operating mode under fault or abnormal conditions to ensure basic system functions;

[0235] Control parameter configuration under different modes: obtained based on multi-objective optimization, and dynamically fine-tuned according to the real-time system status;

[0236] Normal mode: The charging and discharging current is 80% of the rated value, and the temperature rise is allowed to not exceed 15℃, balancing energy efficiency and lifespan;

[0237] High-performance mode: Charge and discharge current is 100% of the rated value, temperature rise is allowed not to exceed 20°C, and response speed is optimized;

[0238] Energy-saving mode: The charging and discharging current is 60% of the rated value, and the temperature rise is allowed to not exceed 10℃ to minimize energy loss;

[0239] Life extension mode: the charge / discharge current is 50% of the rated value, the temperature rise is allowed to not exceed 5°C, and the cyclic stress is minimized;

[0240] Emergency mode: Based on the fault diagnosis results, the optimal safety strategy is adopted to ensure system controllability.

[0241] S400, based on the generated adaptive balance control strategy, achieves precise control of the capacitor bank through the power electronic interface and establishes a closed-loop feedback mechanism to obtain the control execution results and the new system state;

[0242] The specific steps are as follows:

[0243] S410, control command parsing and preprocessing;

[0244] Control signal conversion: The control strategy is converted into a specific PWM control signal, using an adaptive duty cycle mapping algorithm.

[0245]

[0246] PWM duty Indicates the PWM duty cycle, k comp (V i ,T i () represents the voltage-temperature compensation coefficient, obtained through table lookup and linear interpolation; I represents the target current value; I max Indicates the maximum allowable current value; V i T represents the voltage of the i-th capacitor unit; i This represents the temperature of the i-th capacitor cell.

[0247] The smooth control signal generation algorithm uses S-curve trajectory planning to avoid electrical shocks caused by sudden changes in the control signal.

[0248]

[0249] Where I smooth (t) represents the current value after smoothing; I initial Indicates the initial current value; I target The target current value is represented by t0; the start time is represented by t. ramp This indicates the ramp-up time, which is adaptively calculated based on the current change amplitude and temperature.

[0250] Multiple safety constraint checks: Ensure control commands are within the system's safe range by limiting current and voltage ranges and dynamically derating based on temperature and health status.

[0251] S420, multi-layer closed-loop control implementation:

[0252] Enhanced inner-loop control employs an active disturbance rejection current control algorithm with a control period of 5μs.

[0253]

[0254] Where u I (t) represents the control output; K p and K iThese represent the parameters of the first and second PI controllers, with default values ​​of 0.5 and 0.05 respectively; I ref (t) represents the reference current; I(t) represents the actual current; f ESO (t) represents the disturbance compensation amount output by the extended state observer; τ represents the integral variable; t represents the current time.

[0255] Predictive outer-loop control, employing model-based predictive control, has a control period of 500μs and generates a current reference value.

[0256]

[0257] [K pv ,K iv ,K dv ] = f adapt (SOC i U total Mode)

[0258] Where I ref (t) represents the reference current; K pv K iv and K dv These represent the second, third, and fourth control parameters of the outer-loop PID controller, respectively, determined by the adaptive function f. adapt Based on the state of charge (SOC) of the capacitor unit i Unbalance U total Dynamic adjustment of control mode; V ref (t) represents the reference voltage; V(t) represents the actual voltage; I ref (t) represents the generated current reference value; τ represents the integration variable; t represents the current time; This represents the time derivative operator.

[0259] Cross-coupling suppression control addresses the mutual interference between multiple capacitor units:

[0260]

[0261] in This represents the current reference value after adding cross-coupling compensation; I ref,i (t) represents the original current reference value; γ ij The coupling coefficient matrix of unit i to unit j is obtained through online system identification; V j (t) and V ref,j (t) represents the actual voltage and reference voltage of cell j, respectively; n represents the total number of capacitor cells; i and j represent the capacitor cell indices; t represents the current time.

[0262] S430, achieved through high-efficiency power conversion technology;

[0263] It adopts an innovative bidirectional DC-DC converter topology and supports 0-5A continuously adjustable charging and discharging current.

[0264] High-performance silicon carbide metal-oxide-semiconductor field-effect transistor (SiC MOSFET) driving technology:

[0265] Switching frequency: dynamically adjusted according to energy efficiency and noise, ranging from 50-100kHz;

[0266] Adaptive dead-time control, dynamically adjusted according to the rate of change of current, with a basic dead-time of 150ns;

[0267] Optimize gate drive voltage: +20 / -5V to suppress false triggering;

[0268] Fast switching technology: rise / fall time <50ns, reducing switching losses;

[0269] Innovative soft-switching technology is implemented:

[0270] Zero-voltage turn-on (ZVS): Achieved using a resonant network to reduce turn-on losses;

[0271] Zero current turn-off (ZCS): Achieved using phase shift control triggered by current sensing;

[0272] Adaptive resonance parameter optimization: The values ​​of resonant inductance and capacitance are dynamically adjusted according to the load current to ensure optimal soft-switching conditions;

[0273] Electromagnetic compatibility design: Multi-stage common-mode and differential-mode filters are added to suppress electromagnetic interference. Key parameters include:

[0274] Common mode inductance: 5mH, saturation current 5A;

[0275] Y capacitor: 2.2nF / 3kV, safety certification level;

[0276] X capacitor: 1μF / 450V, high ripple current capability;

[0277] Differential mode inductor: 10μH, low DCR design;

[0278] S440, Intelligent Thermal Management and Control System;

[0279] Innovative temperature equalization strategy: Temperature prediction control based on neural networks.

[0280]

[0281] in Indicates the predicted future temperature; f NN T represents a trained neural network prediction model;i (t), I i (t) and V i (t) represents the current temperature, current, and voltage, respectively; T amb The ambient temperature is represented; Δt represents the prediction time step, set to 30 seconds; and i represents the capacitor cell index.

[0282] Temperature prediction-based control strategy:

[0283] For high-temperature capacitor cells (predicted temperature higher than average temperature plus a threshold), reduce the charging current and increase the discharging current.

[0284] For low-temperature capacitor cells (predicted temperature is below the average temperature minus the threshold), increase the charging current and decrease the discharging current.

[0285] The current adjustment coefficient is dynamically calculated based on the temperature difference; the larger the temperature difference, the more noticeable the adjustment.

[0286] Innovative Active Thermal Runaway Prevention System: When a risk of thermal runaway is detected (temperature exceeds the warning threshold or the rate of temperature rise is abnormal), a three-level gradient response mechanism is activated:

[0287] Level 1 (Minor Risk): Fan speed increased to 70%, charging current reduced to 80%, soft-start thermal protection mode activated;

[0288] Level 2 (Medium Risk): Fans run at full speed, charging current decreases to 50%, discharging current increases to 120%, and auxiliary cooling system is activated;

[0289] Level 3 (Severe Risk): Immediately disconnect the charging path, rapidly release energy through the maximum permissible discharge current, activate the emergency cooling system, and trigger a system alarm.

[0290] S450, multi-dimensional performance evaluation and feedback;

[0291] Real-time control effect evaluation system:

[0292] Calculate the absolute change and relative percentage change in imbalance before and after control to evaluate the control effect;

[0293] Comprehensive control efficiency evaluation model:

[0294]

[0295] Where E consumed The energy consumed by the control process is expressed in joules; |ΔSOH| represents the loss of health status caused by the control process; α soh This represents the weighting coefficient for health status, with a default value of 2.0; η control This represents a control efficiency index that considers both energy and health factors; |ΔUtotal | represents the absolute value of the change in imbalance.

[0296] An innovative multi-dimensional feedback mechanism is used to construct structured feedback data packets, which are then passed to the S3 learning algorithm to update the reinforcement learning system.

[0297] Feedback={State,Action,Reward,NextState,Metrics}

[0298] Where State represents the system state before control; Action represents the control action performed; Reward represents the reward value obtained; NextState represents the system state after control; Metrics represents additional evaluation metrics; the reward value calculation comprehensively considers the improvement of imbalance, control efficiency, improvement of temperature balance, and the impact of health status. Additional evaluation metrics include imbalance change, energy index, temperature index, time index, health index, and comprehensive performance index.

[0299] S500 assesses the health status of capacitor banks, predicts remaining service life, and develops predictive maintenance strategies based on new system status and historical operating data.

[0300] Based on the new system status and acquired historical operating data from the S400, the health status of the capacitor bank is assessed, its remaining service life is predicted, and predictive maintenance strategies are developed. This embodiment innovatively employs a combination of multi-source data fusion and deep learning algorithms to achieve high-precision health status assessment and service life prediction.

[0301] S510, multi-index capacitor cell health status (SOH) assessment;

[0302] Multi-dimensional capacity decay assessment model:

[0303]

[0304] Where SOH C Represents a health status index based on capacity; C current Indicates the current capacity; C rated Indicates rated capacity; f C This indicates that the number of iterations n is considered. cycle Depth of discharge (DOD) and average temperature (T) avg The compensation function;

[0305] Advanced internal resistance growth assessment model:

[0306]

[0307] Where SOH R R represents a health status index based on internal resistance.initial R represents the initial internal resistance; critical R represents the critical internal resistance (default is twice the initial value); current Indicates the current internal resistance; g R This indicates that the average current I is considered. avg and temperature change T var The correction function;

[0308] Self-discharge characteristic evaluation indicators:

[0309]

[0310] Where α initial Indicates the initial self-discharge rate; α threshold Indicates the critical self-discharge rate (default is 5 times the initial value); α current Indicates the current self-discharge rate; SOH α This represents a health status index based on self-discharge rate.

[0311] Temperature response characteristics assessment:

[0312]

[0313] Where τ initial τ represents the initial temperature response time constant; current Represents the current temperature response time constant; γ represents the temperature response characteristic weighting coefficient, with a value of 0.8; SOH T It represents a health status index based on temperature response characteristics.

[0314] Innovative multidimensional integrated health index:

[0315]

[0316] Where SOH i SOH C SOH R SOH α and SOH T ;w i (t) represents the dynamic weighting coefficient, which is adaptively adjusted based on current operating conditions and historical data:

[0317]

[0318] Where w i,base This represents the basic weights, with default values ​​of [050, 030, 010, 0.1]; δ i This represents the magnification factor, with a value of 2.0; ρ i (t), ρ j(t) represents the correlation coefficients of the i-th and j-th indicators relative to historical data, respectively; t represents the current time; exp represents the exponential function.

[0319] S520, an innovative residual-enhanced deep learning lifetime prediction technology:

[0320] Deep learning framework: A hybrid deep learning architecture based on LSTM and residual networks, including:

[0321] Input layer: Multi-scale sliding window feature extraction;

[0322] Feature processing layer: Feature selection enhanced with attention mechanism;

[0323] Dual-path LSTM layer:

[0324] First approach: Fast state change capture, 2-layer LSTM, 128 neurons per layer;

[0325] Second approach: Long-term trend analysis, 1-layer LSTM, 256 neurons;

[0326] Residual connection layer, fusing original features with deep features:

[0327] h res =h LSTM +W res ·x processed

[0328] Where h LSTM W represents the output features of the LSTM layer. res The weight matrix represents the residual connection; x processed Represents the processed input features; h res This indicates the characteristics after fusion.

[0329] The fully connected layer employs dropout and batch normalization to enhance generalization capabilities.

[0330] First layer: 256 input nodes, 128 output nodes, ReLU activation, dropout rate 0.3;

[0331] Second layer: 128 input nodes, 64 output nodes, ReLU activation, dropout rate 0.2;

[0332] The third layer has 64 input nodes and 1 output node, with linear activation.

[0333] Innovative multi-source input feature set:

[0334] X = {X static ,X temporal ,X derived ,X external}

[0335] Where X static Indicates static characteristics (such as rated capacity, initial internal resistance, etc.); X temporal Represents time-series characteristics (such as trends in health status, changes in internal resistance, etc.); X derived Indicates derived characteristics (such as average depth of discharge, cumulative energy throughput, etc.); X external Indicates external factors (such as ambient temperature, humidity, etc.);

[0336] Loss function design, combining prediction accuracy and uncertainty assessment:

[0337]

[0338] in Indicates the total loss; Indicates the mean square error loss; Indicates quantile loss; λ1, λ2, and λ3 represent the mean square error loss, quantile loss, and uncertainty loss weighting coefficients, respectively.

[0339] Final prediction output:

[0340] RUL = f LSTM-ResNet (X)±σ pred

[0341] Where f LSTM-ResNet This represents the trained hybrid deep learning model; RUL represents the predicted remaining lifetime in hours; σ pred It represents the uncertainty of the prediction and reflects the confidence interval of the prediction result.

[0342] S530, Advanced Fault Mode Recognition and Early Warning;

[0343] Construction of a multi-source failure mode library: Based on historical data and expert knowledge, a knowledge base containing the following typical failure modes will be constructed:

[0344] Capacity decay type: characterized by capacity and its rate of change characteristic curve;

[0345] Internal resistance growth type: characterized by the characteristic curve of internal resistance and its rate of change;

[0346] Self-discharge increasing type: characterized by self-discharge rate and its variation characteristic curve;

[0347] Temperature anomaly type: characterized by temperature and temperature gradient characteristic curves;

[0348] Deteriorating balance performance type: characterized by the degree of imbalance and its change characteristic curve;

[0349] Multi-class diagnostic system: Employs an ensemble learning framework to integrate the results of multiple classifiers;

[0350] Cascaded Random Forest (RF): for initial feature selection;

[0351] Support Vector Machine (SVM): Core features for accurate classification;

[0352] Gradient Boosting Decision Tree (GBDT): Handles complex feature interactions;

[0353] Fully connected neural networks (DNNs): capture high-dimensional nonlinear relationships;

[0354] Adaptive anomaly detection algorithm: Combining the Local Anomaly Factor (LOF) and the Isolation Forest (IF), the algorithm calculates anomaly scores by weighted fusion of the results of the two algorithms, thereby achieving dual verification of anomaly states.

[0355] Multi-level early warning mechanism: Generates tiered early warnings based on the level of fault risk;

[0356] Level 1 Warning (Alert Level): Abnormal score, remaining life expectancy, or health status reaches a minor warning threshold;

[0357] Level 2 Warning (Attention Level): Abnormal score, remaining life expectancy, or health status reaches the medium warning threshold;

[0358] Level 3 warning (warning level): Anomaly score, remaining lifespan, or health status reaches a critical warning threshold, or the same fault mode characteristic is detected continuously.

[0359] S540, an innovative model integrating health status and balance control strategies;

[0360] Unified mathematical model of multi-objective health status:

[0361] SOH=f(C relative ,R increase ,α rate ,T deviation U balance )

[0362] Where SOH represents the health state; f represents the multi-objective health state function; C relative Represents relative capacity; R increase Indicates the internal resistance growth rate; α rate Indicates the change in self-discharge rate; T deviation Indicates temperature deviation; U balance This indicates the cumulative historical impact of imbalance.

[0363] Simplified representation of the mathematical model:

[0364] SOH = 100-w C ·(1-Crelative )-w R ·R increase -w α ·(α rate -1)-w T

[0365] ·T deviation -w U ·U balance

[0366] Where w C w R w α w T and w U These represent the weighting coefficients for the relative capacity, internal resistance growth rate, self-discharge rate change, and historical cumulative impact of imbalance, respectively, which are determined through neural network training and optimization.

[0367] S550, based on a predictive lifetime-aware balance control strategy;

[0368] Hybrid models based on LSTM and residual networks include:

[0369] Input layer: Multi-dimensional feature input, including historical SOH data, historical temperature data, historical current data, etc.;

[0370] LSTM layer: Extracts temporal features;

[0371] Residual connection: fuses the original features with the output features of the LSTM;

[0372] Fully connected layer: Generates the final prediction result;

[0373] Lifetime balancing control function:

[0374] u i (t)=u base (t)·g(SOH i (t),RUL i (t))

[0375] Where u i (t) represents the control input of the i-th capacitor unit; u base (t) represents the baseline control quantity; g(SOH) i (t),RUL i (t) represents a two-factor adjustment function based on health status and remaining life expectancy:

[0376] g(SOH i (t),RUL i (t))=ω1·g SOH (SOH i (t))+ω2·gRUL (RUL i (t))

[0377] Where ω1 and ω2 represent the weighting coefficients for health status and remaining life expectancy, respectively, with default values ​​of 0.7 and 0.3; g SOH and g RUL These are the adjustment functions based on SOH and RUL, respectively;

[0378] This function ensures that capacitors in poor health or with short remaining lifespan receive more stringent protection, while capacitors in good health bear more of the system load, thereby maximizing the overall lifespan of the system.

[0379] S560, Intelligent Predictive Maintenance Decision Support System;

[0380] The optimal maintenance timing is determined based on a multi-objective lifetime-cost optimization model:

[0381]

[0382] Where T maintenance Indicates the optimal maintenance timing; C maintenance (T) represents the cost function for maintenance at time T; C failure (T) represents the failure risk cost function, which is related to the predicted failure probability; C operation (T) represents the cost of maintenance on system operation at time T; w main w fail and w oper These represent the weighting coefficients for maintenance costs, risk costs, and operational impact costs, respectively, which are determined by the system priority; argmin represents the value of the independent variable that minimizes the objective function.

[0383] Maintain priority intelligent sorting:

[0384] Itemized risk score calculation: the product of failure probability, failure impact severity, and maintenance urgency;

[0385] Failure probability P i Calculation: A combination of a linear mapping based on remaining useful life and the base failure probability;

[0386] Fault severity S i Based on system architecture and unit importance assessment, it is divided into 3 levels:

[0387] Critical level (S) i =3): The core unit that affects the stable operation of the entire system;

[0388] Importance level (S) i =2): Key units that affect some functions;

[0389] General level (S) i =1): Non-critical units in redundant design;

[0390] Maintenance urgency: determined based on the rate of change in health status;

[0391] Group optimization and maintenance strategy:

[0392] Clustering algorithms are used to group units with similar remaining useful lives.

[0393] Intra-team collaborative maintenance planning;

[0394] Maintenance interval optimization: Ensure that maintenance opportunities are distributed reasonably;

[0395] Maintaining decision output and execution:

[0396] Based on the unit's health status, fault type, and system operation requirements, the system generates the following types of maintenance recommendations:

[0397] Replacement recommendation: Units with low health status, serious fault characteristics, or short remaining life;

[0398] Balance adjustment recommendation: Capacitor banks with large imbalance;

[0399] Parameter recalibration recommendations: Units whose health status assessment does not match their actual performance;

[0400] Deep balancing recommendation: Capacitor banks with continuously increasing imbalance;

[0401] Preventative inspection recommendations: Units approaching maintenance thresholds;

[0402] Maintenance execution feedback: After performing maintenance, the system collects execution effect data and updates the health status assessment and prediction model.

[0403] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for intelligent capacitor bank balance optimization control, characterized in that, Includes the following steps: Real-time parameters of each unit in the capacitor bank are collected through a distributed sensor network to establish a dynamic data model of the capacitor bank. Based on the real-time parameters of each unit of the capacitor bank collected, a multi-dimensional imbalance assessment model is constructed to quantify the imbalance state of the capacitor bank. Based on the quantification results of the unbalanced state of the capacitor bank, an adaptive balance control strategy is generated by combining deep reinforcement learning and two-layer closed-loop control. Based on the generated adaptive balance control strategy, the capacitor bank is precisely controlled through the power electronic interface, and a closed-loop feedback mechanism is established to obtain the control execution results and the new system state. Based on the new system status and historical operating data, assess the health status of the capacitor bank, predict its remaining service life, and develop predictive maintenance strategies.

2. The intelligent capacitor bank balance optimization control method according to claim 1, characterized in that, Real-time parameters of each unit in the capacitor bank are collected through a distributed sensor network to establish a dynamic data model of the capacitor bank, including: For each capacitor cell in the capacitor bank High-precision sensors are used to collect voltage, current, surface temperature and internal resistance values, where n represents the number of capacitor units; An adaptive sampling frequency algorithm based on capacitance characteristics is adopted to automatically adjust the sampling frequency according to the state of the capacitor bank. The sampling frequency is a weighted combination of the base sampling frequency and the standard deviations of voltage, current and temperature. Based on the collected basic parameters, characteristic parameters are calculated, including effective capacity, self-discharge rate, temperature gradient, and internal resistance change rate. A multi-parameter dynamic correlation model is constructed, and the early identification of multi-parameter abnormal states is achieved by calculating the dynamic correlation matrix between parameters.

3. The intelligent capacitor bank balance optimization control method according to claim 1, characterized in that, Based on the real-time parameters of each unit in the capacitor bank collected, a multi-dimensional imbalance assessment model is constructed to quantify the imbalance state of the capacitor bank, including: Calculate single-parameter imbalance, including voltage imbalance, temperature imbalance, capacity imbalance, and internal resistance imbalance; The parameter weights are adaptively determined based on the entropy weight method by calculating the information entropy, difference coefficient and weight of the parameters. The overall imbalance degree is calculated based on the adaptive weight adjustment of the working conditions. The overall imbalance degree is the weighted sum of the imbalance degrees of each individual parameter, and the weights are dynamically adjusted according to the working conditions. An improved fuzzy C-means clustering algorithm is used to classify the imbalance state into voltage imbalance, temperature imbalance, capacity imbalance, internal resistance imbalance and mixed type. The trend of imbalance is calculated using the adaptive exponential weighted moving average method, and the future trend of imbalance is predicted using the ARIMA model.

4. The intelligent capacitor bank balance optimization control method according to claim 1, characterized in that, Based on the quantization results of the capacitor bank's unbalanced state, and combined with deep reinforcement learning and two-layer closed-loop control, an adaptive balance control strategy is generated, including: A two-layer control architecture is constructed. The inner layer control is based on an enhanced PID controller to achieve millisecond-level fast response. The PID controller includes proportional, integral, derivative and nonlinear correction terms. The outer control layer achieves long-term optimized scheduling based on a fusion of deep reinforcement learning and MPC control; An improved deep deterministic strategy gradient algorithm is adopted, including a multi-scale time-aware Actor network structure and a two-stream Critic network structure; Design a dynamic adaptive fusion mechanism for deep reinforcement learning and model predictive control, and dynamically adjust the weights of the two control strategies through adaptive fusion coefficients; A recursive least squares method combining particle filtering is used to achieve high-precision identification of time-varying system parameters.

5. The intelligent capacitor bank balance optimization control method according to claim 4, characterized in that, Improved deep deterministic policy gradient algorithms include: The state space is defined as a vector containing voltage, current, temperature, internal resistance, state of charge, and unbalance parameters; The action space is defined as a vector containing charging current, discharging current, charging time, discharging time, and control mode. The multi-objective reward function is designed as a weighted combination to reduce overall imbalance, energy loss and maximum temperature rise, while improving energy efficiency and reducing health status loss. An improved priority experience replay mechanism is adopted, and the priority calculation takes into account the impact of TD error and the number of times the sample is sampled.

6. The intelligent capacitor bank balance optimization control method according to claim 1, characterized in that, Based on the generated adaptive balance control strategy, precise control of the capacitor bank is achieved through a power electronic interface, including: The control strategy is converted into a PWM control signal, and an adaptive duty cycle mapping algorithm is adopted. The duty cycle is related to the current value, voltage and temperature. S-curve trajectory planning is used to generate smooth control signals, avoiding electrical shocks caused by sudden changes in control signals; Multi-layer closed-loop control is achieved. The inner loop control adopts an active disturbance rejection current control algorithm with a control period of 5μs, while the outer loop control adopts model-based predictive control with a control period of 500μs. It adopts a bidirectional DC-DC converter topology, supports 0-5A continuously adjustable charging and discharging current, and combines soft-switching technology with zero voltage turn-on and zero current turn-off. Temperature prediction control based on neural networks enables intelligent thermal management, including an active thermal runaway prevention system with gradient response.

7. The intelligent capacitor bank balance optimization control method according to claim 1, characterized in that, Based on the new system status and historical operating data, assess the health status of the capacitor bank and predict its remaining service life, including: Calculate multi-dimensional health status indicators, including health status indices based on capacity, internal resistance, self-discharge rate, and temperature response characteristics; A multi-dimensional health index is integrated using dynamic weighting coefficients, which are adaptively adjusted based on current working conditions and historical data. The remaining useful life is predicted based on a hybrid deep learning architecture of LSTM and residual network. The input feature set includes static features, temporal features, derived features and external factors. A comprehensive loss function combining mean square error, quantile loss, and uncertainty loss is adopted; The final predicted output is the output value of the residual network model and its uncertainty.

8. The intelligent capacitor bank balance optimization control method according to claim 7, characterized in that, Also includes: A multi-source failure mode library is constructed, including failure modes such as capacity decay, internal resistance increase, self-discharge increase, temperature anomaly, and balance performance deterioration. A multi-class diagnostic system employing an ensemble learning framework integrates classification results from cascaded random forests, support vector machines, gradient boosting decision trees, and fully connected neural networks. An adaptive anomaly detection algorithm combining local anomaly factors and isolated forests is used to calculate anomaly scores by weighted fusion of the results of the two algorithms. A three-level gradient early warning mechanism is generated based on the fault risk level; Based on the predictive lifetime-aware balance control strategy, the control input is calculated as a function product of the unit's baseline control quantity, health status, and remaining lifetime, ensuring that capacitor units with poor health status or short remaining lifetime are more strictly protected.

9. The intelligent capacitor bank balance optimization control method according to claim 1, characterized in that, Developing predictive maintenance strategies includes: The optimal maintenance timing is determined based on a multi-objective lifetime-cost optimization model, which is achieved by minimizing the weighted sum of maintenance cost, failure risk cost, and operational impact cost. Calculate maintenance priority; the risk score is the product of failure probability, failure severity, and maintenance urgency. Based on clustering algorithms, cells with similar remaining lifespans are grouped together, and a group optimization maintenance strategy is implemented. Based on the unit's health status, fault type, and system operation requirements, replacement suggestions, balancing adjustment suggestions, parameter recalibration suggestions, deep balancing suggestions, and preventive detection suggestions are generated. After maintenance is performed, execution effect data is collected, and the health status assessment and prediction models are updated.

10. A smart capacitor bank balancing optimization control system, used to execute the steps in the smart capacitor bank balancing optimization control method as described in any one of claims 1-9, characterized in that, include: A distributed sensor network is used to collect the voltage, current, surface temperature, and internal resistance of each unit in the capacitor bank. The edge layer computing unit is used for data processing, feature parameter calculation and imbalance evaluation. It uses the time series database InfluxDB to store data and sets up a hierarchical caching mechanism. The cloud-based intelligent decision-making center includes a deep reinforcement learning module and a model prediction control module, used to generate adaptive equilibrium control strategies; The power electronic interface adopts a bidirectional DC-DC converter topology and integrates SiCMOSFET drive technology and soft-switching technology. A closed-loop feedback controller enables multi-layer closed-loop control, including inner-loop active disturbance rejection current control and outer-loop predictive control. The health status assessment and prediction module, based on a hybrid deep learning architecture of LSTM and residual network, realizes health status assessment and remaining useful life prediction. The predictive maintenance decision support module is used for fault mode identification, anomaly detection, early warning generation, and maintenance strategy formulation.

Images