Electric automatic control method and system based on artificial intelligence

Abstract

The invention relates to the electrical field, and particularly discloses an electrical automatic control method and system based on artificial intelligence, and the method comprises the steps: constructing a load characteristic fingerprint through non-intrusive monitoring; generating a control adaptation parameter set by using the deep analysis model; reconstructing a control logic topology based on the parameter set; performing online incremental learning optimization on the model according to actual operation data; the system comprises a non-intrusive monitoring module, a load characteristic analysis module, a control logic reconstruction module and an online learning optimization module. The technical bottleneck of parameter solidification of a traditional control system is broken through, fundamental transformation from passive response to active adaptation is achieved, through a complete closed loop of'monitoring-analysis-control-optimization ', the control system has the self-evolution ability, the long-standing hidden load compatibility problem is effectively solved, and the control system has the advantages of being simple in structure and convenient to operate. The system compatibility, the operation energy efficiency and the equipment service life are obviously improved, and an innovative intelligent solution is provided for the field of electrical automation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electrical technology, in particular to an electrical automatic control method and system based on artificial intelligence. BACKGROUND

[0002] Electrical automatic control technology has a wide range of applications in industrial automation, smart grid, building electrical and other fields. The current mainstream electrical automatic control system mainly relies on preset control parameters and fixed control logic. This "static" control method exposes serious technical limitations when facing the dynamic changes of load devices in actual operation.

[0003] The main problems existing in the prior art are reflected in the following aspects: first, the traditional control system has a serious lack of understanding of the characteristics of the load. The control system is usually designed based on the rated parameters and idealized model of the load, while the load devices in actual operation often exhibit complex nonlinear and time-varying characteristics. Especially when the system is connected to non-standard devices, old devices or devices from different manufacturers, due to the lack of accurate perception of the real characteristics of the load, the control system can only use a general conservative strategy, resulting in the system always running in a "non-optimal" state.

[0004] Secondly, the existing control strategy lacks self-adaptation. When the load characteristics change slowly due to device aging, environmental changes or changes in operating mode, the traditional control system cannot automatically adjust the control parameters, resulting in gradual degradation of control performance. This performance degradation is often hidden and difficult to detect before failure, but it can significantly affect system efficiency and device life. For example, the traditional control system cannot effectively cope with problems such as changes in characteristics of motor loads due to bearing wear, parameter drift of power electronic devices due to component aging, etc.

[0005] The deeper problem is that the existing technology lacks a complete technical chain from load characteristic perception to control strategy optimization. Although there have been attempts to introduce artificial intelligence technology for device state monitoring or fault warning, most of these solutions remain at the "perception" level and fail to achieve deep integration with the control system. Some load identification technologies can only complete device classification, but cannot convert the identification results into specific control parameter optimization recommendations; some adaptive control methods can adjust control parameters, but lack deep analysis capabilities for load characteristics, making it difficult to achieve precise matching.

[0006] In terms of control system architecture design, the existing solutions generally have the phenomenon of "information island". The load monitoring, characteristic analysis and control execution are independent of each other, and lack effective closed-loop optimization mechanisms. The operation data obtained by the control system cannot be fed back to the analysis model, resulting in the system lacking the ability to continuously learn and evolve. This architectural defect makes it difficult for the control system to adapt to various changes in long-term operation and to form effective knowledge accumulation and experience inheritance.

[0007] Therefore, a new electrical automatic control method and system capable of deeply integrating load characteristic perception, intelligent analysis, control reconstruction and online optimization are urgently needed to fundamentally solve the adaptability problem of traditional control systems when facing dynamic changing loads and improve the intelligent level and long-term operation performance of the system. SUMMARY

[0008] The purpose of the present application is to solve the problems existing in the prior art and to propose an electrical automatic control method and system based on artificial intelligence.

[0009] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0010] The electrical automatic control method based on artificial intelligence comprises:

[0011] Obtaining real-time dynamic response characteristics of load devices in the electrical system through non-invasive monitoring to construct a load characteristic fingerprint;

[0012] Inputting the load characteristic fingerprint into a load characteristic deep analysis model to output a control adaptation parameter set, the control adaptation parameter set containing key parameters that need to be adjusted by the control system and the optimization direction thereof;

[0013] Based on the control adaptation parameter set, reconstructing the control logic topology of the electrical control system to realize dynamic matching of the control system and the load characteristics;

[0014] Based on the actual operation data of the reconstructed control system, generating a control efficiency evaluation index, and performing online incremental learning optimization on the load characteristic deep analysis model according to the control efficiency evaluation index.

[0015] As a further technical solution of the present application, the real-time dynamic response characteristics of load devices in the electrical system are obtained through non-invasive monitoring to construct a load characteristic fingerprint, specifically comprising:

[0016] Applying a pseudo-random binary sequence test signal with an amplitude not exceeding ±2% of the rated value at the power inlet of the electrical system, synchronously collecting the voltage response and current response of the load device;

[0017] Based on the voltage response and current response, extracting four types of real-time dynamic response characteristics;

[0018] Normalizing and feature encoding the four types of real-time dynamic response characteristics to generate a load characteristic fingerprint with uniqueness, the load characteristic fingerprint adopting a three-level encoding structure, including a basic electrical parameter layer, a dynamic response characteristic layer and a running mode characteristic layer.

[0019] As a further technical solution of the present application, the four types of real-time dynamic response characteristics specifically comprise:

[0020] Impedance frequency characteristics, impedance amplitude and phase angle of 10 characteristic frequencies in the frequency range of 0.1Hz to 2kHz are measured;

[0021] Starting transient process, capture the current waveform of the first 100ms at the load starting moment with a sampling rate of 100kHz, extract the starting current peak, rise time constant and oscillation damping ratio;

[0022] Harmonic response characteristics, measure the response amplitude and phase shift of the load to 3rd, 5th and 7th characteristic harmonics;

[0023] Power dynamic characteristics, record the power factor change trajectory and response time of the load during the step change from no load to full load.

[0024] As a further technical solution of the application, the load characteristic deep analysis model adopts a five-layer hybrid neural network structure, comprising:

[0025] The input adaptation layer receives the four types of feature vectors of the load characteristic fingerprint, and respectively performs normalization and feature enhancement through four special feature preprocessing channels;

[0026] The feature fusion layer deeply fuses the four types of preprocessed features through cross-attention mechanism to generate a unified multi-dimensional feature representation;

[0027] The time sequence memory layer adopts a bidirectional long short-term memory network to capture the time evolution law and dynamic behavior pattern of the load characteristics;

[0028] The analysis and reasoning layer analyzes the identity features, health status and adaptation requirements of the load through three parallel fully connected network branches respectively;

[0029] The output mapping layer maps the analysis results into a specific control adaptation parameter set, including PID parameter optimization suggestions, filter parameter configuration, protection threshold setting and optimization direction indication;

[0030] The load characteristic deep analysis model is trained end-to-end through the corresponding relationship between historical load characteristic data and optimal control parameters, and a multi-task learning strategy is adopted during the training process to simultaneously optimize three objective functions of load identification accuracy, health state evaluation accuracy and control parameter recommendation effect.

[0031] As a further technical solution of the application, the load characteristic fingerprint is input into the load characteristic deep analysis model, and a control adaptation parameter set is output, specifically including:

[0032] The three-level encoding structure features of the load characteristic fingerprint are split into four feature vectors, which are input into the input adaptation layer, wherein the basic electrical parameter layer features are input into the impedance characteristic channel, the dynamic response feature layer features are input into the transient response channel, the operating mode feature layer features are input into the behavior mode channel, and the historical operation data are input into the evolution trend channel;

[0033] In the feature fusion layer, the correlation weights between the four types of feature vectors are calculated through the cross-attention mechanism to generate a fusion feature representation with context awareness, wherein the correlation weights of the impedance characteristics and the transient response are used to represent the stability features of the load, and the correlation weights of the behavior mode and the evolution trend are used to represent the adaptability features of the load.

[0034] The time sequence memory layer receives the fused feature sequence, extracts the forward and backward time dependence through a bidirectional long short-term memory network, and outputs the dynamic evolution trajectory of the load characteristics, including the load aging trend, the performance decay rate and the environmental adaptability.

[0035] The three parallel branches of the analysis reasoning layer process the dynamic evolution trajectory respectively, wherein: the identity recognition branch outputs the specific type and manufacturer features of the load device through a softmax classifier; the health assessment branch outputs the residual life prediction and performance decay index of the load through a regression network; and the adaptation analysis branch outputs the optimization direction and adjustment amplitude of the control parameters through a deep reinforcement learning network.

[0036] The output mapping layer synthesizes the analysis results of the three branches to generate a control adaptation parameter set containing the following contents: PID parameter optimization suggestions, including the adjustment amplitude of the proportional coefficient, the optimization direction of the integral time constant and the correction amount of the differential coefficient; filter parameter configuration, including the cutoff frequency setting of the low-pass filter, the center frequency configuration of the notch filter and the filter order selection; protection threshold setting, including the action value and time limit of the overcurrent protection, the startup threshold of the overvoltage protection and the recovery hysteresis of the undervoltage protection; optimization direction indication, providing adjustment priority ranking and adjustment step size suggestion for each parameter.

[0037] The control adaptation parameter set further contains the confidence evaluation of each parameter and the recommended implementation time window, and when the confidence is lower than the preset threshold, the system adopts a gradual adjustment strategy to implement parameter optimization in stages.

[0038] As a further technical solution of the present application, based on the control adaptation parameter set, the control logic topology of the electrical control system is reconstructed to realize the dynamic matching of the control system and the load characteristics, specifically including:

[0039] According to the PID parameter optimization suggestion in the control adaptation parameter set, the core algorithm structure of the controller is reconstructed, including: based on the adjustment amplitude of the proportional coefficient, reconfiguring the gain weight of the feedforward control channel; based on the optimization direction of the integral time constant, reconstructing the anti-saturation mechanism and reset logic of the integral term; based on the correction amount of the differential coefficient, reconstructing the differential lead control structure and filter compensation network;

[0040] According to the filter parameter configuration, the topology connection of the signal processing link is reconstructed, including: based on the cutoff frequency setting, reconstructing the cascade structure of the multistage low-pass filter; based on the center frequency configuration, reconstructing the resonance point tracking mechanism of the adaptive notch filter; based on the filter order selection, reconstructing the phase compensation network of the signal preprocessing channel;

[0041] According to the protection threshold setting, the decision level of the system protection logic is reconstructed, including: based on the action value and time limit of the overcurrent protection, reconstructing the action sequence and coordination mechanism of the hierarchical protection; based on the starting threshold of the overvoltage protection, reconstructing the feedforward warning and hierarchical response logic of the voltage protection; based on the recovery hysteresis of the undervoltage protection, reconstructing the adaptive judgment condition and timing control of system recovery.

[0042] As a further technical solution of the present application, the process of reconstructing the control logic topology of the electrical control system adopts a modular replacement mechanism, which replaces the corresponding modules in the original control logic with the optimized functional modules by dynamic loading while keeping the basic architecture of the system unchanged, thereby realizing online reconstruction of the control system;

[0043] The reconstructed control logic topology and the load characteristic fingerprint establish a mapping relationship, and when the same or similar load characteristic fingerprint is monitored, the system automatically calls the corresponding optimized control logic topology, thereby realizing accurate matching of the control system and the load characteristics;

[0044] The reconstruction process further includes establishing a version management mechanism of the control logic topology, recording the control parameters, topology structure and running effect of each reconstruction, and forming a self-optimization closed loop of the control strategy.

[0045] As a further technical solution of the present application, based on the actual running data of the reconstructed control system, a control efficiency evaluation index is generated, and the load characteristic deep analysis model is optimized through online incremental learning according to the control efficiency evaluation index, specifically including:

[0046] Real-time running data of the reconstructed control system is collected, including electrical parameter data, control response data and load state data, a control efficiency evaluation index is generated through a multi-dimensional efficiency evaluation algorithm, and the control efficiency evaluation index includes: stability improvement index, energy efficiency optimization index, response acceleration index and reliability enhancement index;

[0047] The control performance evaluation index is input to a model incremental learning engine, running data in a preset time period is retained by using a sliding window mechanism, an online incremental learning optimization is performed on the load characteristic deep analysis model, the online incremental learning optimization process adopts a federated learning framework, knowledge sharing and experience migration among multiple devices are realized on the premise of protecting the data privacy of each subsystem.

[0048] As a further technical solution of the application, the online incremental learning optimization of the load characteristic deep analysis model specifically comprises the following steps:

[0049] The real-time running data is cleaned, aligned and labeled to construct an incremental learning sample set;

[0050] Based on the control performance evaluation result, the contribution weight of each feature in the load characteristic fingerprint is recalculated;

[0051] On the premise of keeping the original model structure unchanged, the gradient descent algorithm is used to perform progressive adjustment on the key parameters of the model, the optimized model parameters are fed back to the load characteristic deep analysis model in real time, a complete closed loop from control execution to model optimization is formed, and the system is ensured to continuously evolve itself;

[0052] The new knowledge obtained by the incremental learning is distilled and fused with the original model knowledge to avoid catastrophic forgetting.

[0053] The electrical automatic control system based on artificial intelligence is used to realize an electrical automatic control method based on artificial intelligence, and comprises:

[0054] A non-invasive monitoring module configured to acquire real-time dynamic response characteristics of load devices in an electrical system and construct a load characteristic fingerprint;

[0055] A load characteristic analysis module configured to input the load characteristic fingerprint to a load characteristic deep analysis model and output a control adaptation parameter set;

[0056] A control logic reconstruction module configured to reconstruct the control logic topology of the electrical control system based on the control adaptation parameter set;

[0057] An online learning optimization module configured to generate a control performance evaluation index based on the actual running data of the reconstructed control system and perform online incremental learning optimization on the load characteristic deep analysis model;

[0058] A data storage and management module configured to store a load characteristic fingerprint library, a control adaptation parameter mapping library and a system running history database.

[0059] The application has the following beneficial effects:

[0060] 1. Achieved the fundamental transformation of the electrical control system from "passive response" and "active adaptation", solved the long-standing implicit load compatibility problem; the load characteristic fingerprint constructed by non-invasive monitoring can accurately identify the unique electrical characteristics of various loads, and then convert the characteristic information into specific control adaptation parameters through the load characteristic deep analysis model, and finally realize the deep matching of the control system and the load characteristics through the control logic topology reconstruction; a complete closed loop from load characteristic perception to control strategy generation is established, which enables the control system to have the intelligent ability of "knowing the load, understanding the load and adapting to the load".

[0061] 2. A "monitoring-analysis-control-optimization" full-link self-evolution system is constructed, which breaks through the technical bottleneck of parameter solidification of traditional control systems; the invention introduces an online incremental learning mechanism based on actual operation efficiency, forming a complete self-optimization closed loop; the control effect is quantified in real time through control efficiency evaluation indicators, and when performance degradation is detected, the incremental learning optimization of the load characteristic deep analysis model is automatically triggered, so that the system can continuously track the changes of the load characteristics and adjust the control strategy accordingly.

[0062] 3. A multi-level knowledge accumulation and transfer learning framework is established, which solves the technical problem of difficult control experience precipitation and reuse in complex industrial site environment; the invention realizes the digitalization of control knowledge and systematic application through the linkage of the three databases of load characteristic fingerprint database, control strategy database and operation effect database; successful control experience can be safely shared between different devices through a federated learning framework, and at the same time, the deep association between load characteristics, control strategies and operation effects is established through a knowledge graph unit, so that the learning achievements of a single device can be quickly replicated to the entire system, and when a new device is connected, the optimized control strategy of similar loads can be directly called, greatly improving the overall intelligent level of the system. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 Flowchart of the electrical automatic control method based on artificial intelligence proposed by the invention;

[0064] Figure 2 Module diagram of the electrical automatic control system based on artificial intelligence proposed by the invention;

[0065] Figure 3 Effect comparison of the method of the invention and the prior art method in Example 1 Figure 1 ;

[0066] Figure 4 Effect comparison of the method of the invention and the prior art method in Example 1 Figure 2 ;

[0067] Figure 5Effect comparison of the method of the present application and the prior art method in Example 1 Figure 3 . DETAILED DESCRIPTION

[0068] In order to make the technical means, creative features, purposes and effects of the present application easy to understand, the present application will be further described below in conjunction with specific embodiments.

[0069] Please refer to the accompanying Figure 1 The electrical automatic control method based on artificial intelligence comprises the following steps:

[0070] S1. Obtain the real-time dynamic response characteristics of the load device in the electrical system through non-invasive monitoring, and construct a load characteristic fingerprint; specifically comprising:

[0071] S11. Apply a pseudo-random binary sequence (PRBS) test signal with an amplitude not exceeding ±2% of the rated value at the power inlet of the electrical system, and synchronously collect the voltage response v(t) and current response i(t) of the load device.

[0072] S12. Based on the voltage response v(t) and current response i(t), extract the following four types of real-time dynamic response characteristics:

[0073] S121. Impedance frequency characteristics, measure the impedance amplitude and phase angle of 10 characteristic frequencies in the frequency range of 0.1 Hz to 2 kHz;

[0074] For each characteristic frequency f k (k=1,2,…,10), calculate the impedance Z k : Where: V k and I k are the Fourier amplitudes of the voltage and current, respectively; the impedance amplitude |Z k | and the phase angle are respectively: Where: Re and Im represent the real part and the imaginary part of the complex impedance Z k , respectively.

[0075] S122. Start-up transient process, capture the current waveform of the first 100 ms at the load start-up moment with a sampling rate of 100 kHz, and extract the start-up current peak I peak , the rise time constant τ and the oscillation damping ratio ζ;

[0076] The start-up current peak I peak is the maximum current amplitude in the first 100 ms, the rise time constant τ is obtained by exponential fitting of the initial rising section, and the oscillation damping ratio ζ is calculated by analyzing the logarithmic decay rate of current oscillation decay: Where: f d is the damping oscillation frequency, and ωn The undamped natural frequency.

[0077] S123. Harmonic response characteristics, measuring the response amplitude and phase shift of the load to the 3rd, 5th, and 7th characteristic harmonics;

[0078] For the hth harmonic (h ∈ {3, 5, 7}), the harmonic response amplitude A h and the phase shift θ h are respectively: Where: V h and are the amplitude and phase of the hth harmonic, and V1 and are the amplitude and phase of the fundamental.

[0079] S124. Power dynamic characteristics, recording the power factor change trajectory PE(t) and response time t r of the load during the step change from no load to full load.

[0080] The power factor PF(t) is defined as: Where: P(t) is the active power, and Q(t) is the reactive power; the response time t r is the time required for the power factor to rise from the no-load value PF0 to 90% of the full-load value PF1.

[0081] S13. Normalize and feature encode the above four types of real-time dynamic response characteristics to generate a unique load characteristic fingerprint. The load characteristic fingerprint adopts a three-level encoding structure, including:

[0082] The basic electrical parameter layer contains the rated voltage V n , rated current I n , rated power P n , and power factor PE r .

[0083] The dynamic response characteristic layer contains the impedance frequency characteristics (|Z k | and ), starting transient parameters (I peak , τ, and ζ), harmonic response characteristics (A h and θ h ), and power dynamic characteristics (PE(t) trajectory and t r ).

[0084] The operating mode characteristic layer contains the steady-state current I ss , steady-state power P ss , and operating time T op under typical load modes.

[0085] Each feature parameter is normalized and encoded into a fixed-length binary string, forming the load characteristic fingerprint.

[0086] S2. Input the load characteristic fingerprint into the load characteristic deep analysis model, and output a control adaptation parameter set containing key parameters that need to be adjusted by the control system and their optimization directions;

[0087] The structure of the load characteristic deep analysis model specifically includes:

[0088] The input adaptation layer receives the four types of feature vectors of the load characteristic fingerprint and performs normalization and feature enhancement through four dedicated channels.

[0089] The feature fusion layer uses a cross-attention mechanism to deeply fuse the preprocessed features and generate a unified multi-dimensional feature representation.

[0090] The time sequence memory layer uses a bidirectional long short-term memory network to capture the time evolution law and dynamic behavior pattern of the load characteristics.

[0091] The analysis and reasoning layer includes three parallel fully connected network branches that analyze the identity features, health status, and adaptation requirements of the load, respectively.

[0092] The output mapping layer maps the analysis results into a specific control adaptation parameter set, including PID parameter optimization suggestions, filter parameter configurations, protection threshold settings, and optimization direction indications.

[0093] The load characteristic deep analysis model is trained end-to-end through the correspondence between historical load characteristic data and optimal control parameters. During the training process, a multi-task learning strategy is adopted to simultaneously optimize three objective functions: load identification accuracy, health status assessment accuracy, and control parameter recommendation effect. Specifically, it includes:

[0094] Data preparation: Collect the correspondence between historical load characteristic data and optimal control parameters to form a training set where f i is the load characteristic fingerprint, p i is the control adaptation parameter set.

[0095] Multi-task learning: The model is trained end-to-end through historical data, simultaneously optimizing three objective functions: load identification accuracy, health status assessment accuracy, and control parameter recommendation effect. A dynamic weight adjustment strategy is used to balance the loss functions of different tasks, avoiding the dominance of a certain task in the training process.

[0096] Loss function: Total objective function is the weighted sum of three sub-task loss functions, i.e. where α, β, γ are dynamically adjusted weight coefficients, Loss functions for load identification, health state assessment and control parameter recommendation, respectively

[0097] Training optimization: minimize the total objective function by optimization algorithms such as gradient descent Constantly adjust model parameters to improve the model's ability to analyze load characteristic data and the effectiveness of control parameter recommendations.

[0098] S2 specifically includes:

[0099] S21. Split the three-level encoding structure features of load characteristic fingerprints into four feature vectors, represented as X base (basic electrical parameter layer), X dynamic (dynamic response feature layer), X mode (operation mode feature layer), and X history (historical operation data), and input them into the corresponding impedance characteristic, transient response, behavior mode, and evolution trend channels.

[0100] S22. In the feature fusion layer, the correlation weights between the four types of feature vectors are calculated through the cross-attention mechanism, as follows:

[0101]

[0102] Where: Q impedance ,K transient ,Q behavior and K evolution are the query vectors and key vectors of impedance characteristics, transient response, behavior mode, and evolution trend, d is the feature dimension, α stability represents the correlation weight between impedance characteristics and transient response, α adaptability represents the correlation weight between behavior mode and evolution trend.

[0103] S23. The time sequence memory layer receives the fused feature sequence X = [x1, x2,..., x T ], extracts the forward and backward time dependence through bidirectional LSTM, and outputs the dynamic evolution trajectory of load characteristics:

[0104]

[0105] Where: x t is the time step feature vector, is the hidden state of the forward LSTM, is the hidden state of the reverse LSTM, h t is the fused hidden state.

[0106] S24. The identity recognition branch outputs the specific type and manufacturer characteristics of the load device through a softmax classifier: where p(y = c|x) represents the probability of the load device belonging to class c given the input feature x, W c and b c are the weight and bias of the softmax classifier, respectively;

[0107] a health assessment branch outputs the remaining life prediction and performance degradation index of the load through a regression network: RUL = W rul · h t + b rul , D = W decay · h t + b decay , where W rul , W decay and b rul , b decay are the weight and bias of the regression network, respectively, RUL represents the remaining useful life, and D represents the performance degradation index;

[0108] an adaptive analysis branch outputs the optimization direction and adjustment amplitude of the control parameters through a deep reinforcement learning network: a * = argmax a Q(h t , a), where Q(h t , a) is the Q value function, a * represents the optimal control action selected by the deep reinforcement learning network, and a represents a possible control action;

[0109] S25. The analysis results of the three branches are integrated to generate a set of control adaptation parameters, including PID parameter optimization suggestions, filter parameter configurations, protection threshold settings, and optimization direction indications, and the confidence C s of the parameters and the time window for recommended implementation are evaluated:

[0110] ΔK p = W pid-p · h t + b pid-p

[0111] ΔT i = sign(W pid-i · h t + b pid-i )

[0112] ΔK d = W pid-d · h t + b pid-d

[0113] C s = σ(W conf · h t + b conf)

[0114] where: ΔK p is the proportional coefficient adjustment range, ΔT i is the integral time optimization direction, ΔK d is the derivative coefficient correction amount, C s is the confidence, W pid-p and b pid-p are the weight matrix and bias of the proportional control parameter respectively, W pid-i and b pid-i are the weight matrix and bias of the integral control parameter respectively, W pid-d and b pid-d are the weight matrix and bias of the derivative control parameter respectively, W conf and b conf are the weight matrix and bias of the confidence evaluation respectively, sign() represents the sign function, and σ() represents the activation function.

[0115] When the confidence C s is lower than the preset threshold, the system implements parameter optimization in stages using a gradual adjustment strategy.

[0116] S3. Reconstruct the control logic topology of the electrical control system based on the control adaptation parameter set to realize dynamic matching of the control system and the load characteristics; specifically including:

[0117] S31. Reconstruct the core algorithm structure of the controller according to the PID parameter optimization suggestion in the control adaptation parameter set;

[0118] S311. Based on the adjustment range of the proportional coefficient, reconfigure the gain weight of the feedforward control channel, and the feedforward control channel gain weight update formula: K ff_new = K ff_old + ΔK p , where: K ff_new is the updated feedforward control channel gain weight, K ff_old is the original feedforward control channel gain weight, and ΔK p is the adjustment range of the proportional coefficient (from the control adaptation parameter set);

[0119] S312. Based on the optimization direction of the integral time constant, reconstruct the anti-saturation mechanism and reset logic of the integral term, and the integral term anti-saturation mechanism: Prevent integral saturation: where: I new is the updated integral term, I old is the current integral term, T i is the integral time constant (the adjustment range of which is determined by the optimization direction), e(t) is the error at the current time, and I max and Imin The maximum and minimum values of the integral term are used to prevent integral saturation;

[0120] S313. Based on the correction amount of the differential coefficient, reconstruct the differential look-ahead control structure and the filter compensation network, the differential look-ahead control formula: D new = K d ·e(t)-e(t-1)), the filter compensation network: Where: D new is the differential control term, K d is the differential coefficient (the correction amount determines its adjustment range), e(t) and e(t-1) are the errors at the current time and the previous time, D filtered is the filtered differential control term, τ is the filter time constant, T d is the differential time constant.

[0121] S32. According to the filter parameter configuration, reconstruct the topology connection of the signal processing link;

[0122] S321. Based on the cutoff frequency setting, reconstruct the cascade structure of the multi-stage low-pass filter, the cutoff frequency is ω c The transfer function of a first-order low-pass filter: The cascade of a multi-stage low-pass filter: Where: H(s) represents the transfer function of the system, ω c is the cutoff frequency, s is the complex frequency variable, H cascade (s) represents the total transfer function after the cascade of the multi-stage system, N is the filter order, is the cutoff frequency of the i-th filter;

[0123] S322. Based on the center frequency configuration, reconstruct the resonance point tracking mechanism of the adaptive notch filter, the transfer function of the notch filter: Where: ω n is the center frequency (resonant frequency), Q is the quality factor (related to the selectivity of the filter);

[0124] S323. Based on the filter order selection, reconstruct the phase compensation network of the signal preprocessing channel, the design of the phase compensation network is based on the group delay characteristics of the filter, to ensure the phase consistency of the signal: Where: τ g (ω) is the group delay, is the phase response of the filter.

[0125] S33. According to the protection threshold setting, reconstruct the decision level of the system protection logic;

[0126] S331. Based on the action value and time limit of the overcurrent protection, reconstruct the action sequence of the hierarchical protection, the overcurrent protection action condition: I(t)>Ith t≥t delay wherein: I(t) is the current current, I th is the overcurrent protection threshold, t is the current time, t delay is the action delay time;

[0127] S332. Reconstruct the feedforward warning and hierarchical response logic of the voltage protection based on the starting threshold of the overvoltage protection, the overvoltage protection warning and response: V(t)>V th , trigger protection action or warning, wherein: V(t) is the current voltage, V th is the overvoltage protection threshold;

[0128] S333. Reconstruct the adaptive judgment condition of system recovery based on the recovery hysteresis of the undervoltage protection, the recovery hysteresis mechanism: V(t)>V th-on , the system recovers and runs, V(t)<V th-off , the system remains stopped, wherein: V th-on is the recovery voltage threshold (upper limit), V th-off is the stop voltage threshold (lower limit).

[0129] The modular replacement mechanism allows dynamic updating of functional modules while keeping the basic architecture of the system unchanged; the establishment of the mapping relationship can be achieved through a pattern recognition algorithm, and when a specific load characteristic fingerprint is detected, the corresponding optimized control logic topology is triggered.

[0130] S4. Based on the actual operation data of the reconstructed control system, generate control efficiency evaluation indexes, and perform online incremental learning optimization on the load characteristic deep analysis model according to the control efficiency evaluation indexes; specifically including:

[0131] S41. Generate control efficiency evaluation indexes through multi-dimensional efficiency evaluation algorithm;

[0132] Stability improvement index: by comparing the system overshoot M p , the regulation time t s and the steady-state error e ss before and after control reconstruction, calculate the stability improvement degree: wherein: I stability is the stability improvement degree; is the weight coefficient, which is determined according to actual needs; the overshoot improvement and are the overshoots before and after; and are the regulation times before and after; the steady-state error improvement and are the front and rear steady-state errors, respectively;

[0133] Energy efficiency optimization index: based on power factor improvement rate ΔPF, harmonic distortion reduction degree ΔTHD and energy loss reduction amount ΔE, comprehensive evaluation of energy efficiency improvement effect: I efficiency = β PF · ΔPF + β THD · ΔTHD + β E · ΔE, wherein: I efficiency is the energy efficiency improvement effect; β PF , β THD , β E are weight coefficients; power factor improvement rate ΔPF = PF after - PF before , PF after and PF before are the front and rear power factors, respectively; harmonic distortion reduction degree ΔTHD = THD before - THD after , THD before and THD after are the front and rear harmonic distortions, respectively; energy loss reduction amount ΔE = E before - E after , E before and E after are the front and rear energy losses, respectively;

[0134] Response acceleration index: by analyzing the response speed t r and the set value tracking accuracy e tr of the system to load changes, the dynamic performance improvement is quantified: wherein: I response is the dynamic performance improvement effect; is the weight coefficient; response time improvement and are the front and rear response speeds, respectively; set value tracking accuracy improvement Δe tr = e tr-before - e tr-aftr , e tr-before and e tr-after are the front and rear set value tracking accuracies, respectively;

[0135] Reliability enhancement index: statistical protection action frequency reduction rate ΔN trip , fault early warning accuracy A fault and equipment operation stability S run , evaluate the system reliability improvement: wherein: I reliability is the system reliability improvement effect; is a weight coefficient; the protection action frequency reduction rate ΔN trip = N trip-before -N trip-after , N trip-before and N trip-after are the front and rear protection action frequencies; A fault is a fault early warning accuracy; S run is a device operation stability;

[0136] S42. The control performance evaluation index is input to the model incremental learning engine, the running data in the recent preset time period is retained by using a sliding window mechanism, and the load characteristic deep analysis model is optimized by online incremental learning;

[0137] The sliding window mechanism is: W = {x(t-N+1), x(t-N+2), …x(t)}, wherein:

[0138] W is a sliding window, x(t) is real-time running data, and N is a window size;

[0139] S421. The real-time running data is cleaned, aligned and labeled, and an incremental learning sample set is constructed;

[0140] S422. Feature importance evaluation: wherein: w i is the contribution weight of the i-th feature, f i is the i-th feature, corr(f i , I efficacy ) is the correlation between the feature f i and the control performance evaluation index I efficacy , and M is the total number of features;

[0141] S423. Model parameter fine-tuning: wherein: θ new is the updated model parameter, θ old is the original model parameter, η is the learning rate, is the gradient of the loss function to the model parameter;

[0142] S424. Knowledge distillation fusion: L total = L original + λ·L distill , wherein: L total is a total loss function, L original is an original model loss, L distill is a distillation loss, and λ is a distillation loss weight.

[0143] Please refer to the attached Figure 2 , the artificial intelligence-based electrical automatic control system, characterized in that it is used to realize an artificial intelligence-based electrical automatic control method, comprising:

[0144] The non-invasive monitoring module is configured to obtain real-time dynamic response characteristics of the load device in the electrical system and construct a load characteristic fingerprint. Specifically, the non-invasive monitoring module comprises:

[0145] The test signal generation unit is configured to generate a pseudo-random binary sequence test signal with an amplitude not exceeding ±2% of the rated value. The test signal generation unit is implemented by using a digital signal processor and can simultaneously generate test signals of multiple frequencies. The test signals are coupled to the electrical system through a power amplifier.

[0146] The multi-channel acquisition unit is configured to synchronously acquire voltage response, current response, and power response of the load device.

[0147] The feature extraction unit is configured to extract impedance frequency characteristics, start transient process, harmonic response characteristics, and power dynamic characteristics from the acquired data.

[0148] The fingerprint construction unit is configured to normalize and encode the extracted features to generate a load characteristic fingerprint.

[0149] The load characteristic analysis module is configured to input the load characteristic fingerprint into a load characteristic deep analysis model and output a control adaptation parameter set. Specifically, the load characteristic analysis module comprises:

[0150] The model inference engine is configured to perform forward calculation of the load characteristic deep analysis model. The model inference engine adopts a heterogeneous computing architecture, in which feature preprocessing and simple inference are performed on an edge computing node, and complex model calculation is performed on a cloud server.

[0151] The feature preprocessing unit is configured to normalize and enhance the input features.

[0152] The multi-task output unit is configured to output load identity recognition, health state evaluation, and control adaptation analysis results in parallel.

[0153] The confidence assessment unit is configured to assess the reliability of the analysis results and recommend the implementation priority.

[0154] The control logic reconstruction module is configured to reconstruct the control logic topology of the electrical control system based on the control adaptation parameter set. Specifically, the control logic reconstruction module comprises:

[0155] The PID parameter reconstruction unit is configured to reconfigure the proportional, integral, and derivative parameters of the controller according to optimization suggestions. The PID parameter reconstruction unit adopts a hot plug mechanism and can seamlessly switch different control algorithms during system operation to ensure the continuity of the control process.

[0156] The filter topology reconstruction unit is configured to reconstruct the filter structure and parameters of the signal processing link.

[0157] A protection logic reconfiguration unit configured to reset the action threshold and timing logic of system protection;

[0158] A version management unit configured to record the change history and running effect of the control logic topology.

[0159] An online learning optimization module configured to generate control performance evaluation indicators based on the actual running data of the reconfigured control system, and perform online incremental learning optimization on the load characteristic deep analysis model; specifically including:

[0160] An efficiency evaluation unit configured to calculate stability improvement indicators, energy efficiency optimization indicators, response acceleration indicators, and reliability enhancement indicators;

[0161] An incremental learning engine configured to fine-tune the parameters of the load characteristic deep analysis model based on real-time running data;

[0162] A federated learning coordinator configured to coordinate knowledge sharing among multiple devices while protecting data privacy;

[0163] A model re-calibration unit configured to trigger model retraining when control performance is below a threshold.

[0164] A data storage and management module configured to store a load characteristic fingerprint library, a control adaptation parameter mapping library, and a system running history database; specifically including:

[0165] A fingerprint feature database configured to store the characteristic fingerprints and evolution history of different load devices;

[0166] A control strategy library configured to store verified effective control logic topologies and parameter settings;

[0167] A running effect library configured to record the actual running performance data of each control strategy;

[0168] A knowledge graph unit configured to establish the association between load characteristics, control strategies, and running effects.

[0169] Further including:

[0170] A real-time communication bus configured to transmit monitoring data, control instructions, and learning parameters among modules;

[0171] A security protection unit configured to ensure system communication security and control instruction integrity;

[0172] A human-computer interaction interface configured to display system status and provide a manual intervention channel;

[0173] A fault emergency unit configured to switch to a backup control mode when the system is abnormal.

[0174] Example 1

[0175] The AI-based electrical automatic control method and system described in the present application were tested in a real-world power distribution system in an industrial park. The test environment included a main power distribution cabinet and multiple load devices, such as induction motors (5.5 kW), servo drives (3 kW), variable frequency drives (7.5 kW), and nonlinear loads (such as rectifiers). These devices came from different manufacturers and had been in use for 1 to 10 years to simulate the diversity of loads in real industrial scenarios.

[0176] Test setup:

[0177] The method of the present application: Deploy the overall solution, including non-intrusive monitoring module, load characteristic analysis module, control logic reconstruction module and online learning optimization module. The system uses a five-layer hybrid neural network model to analyze load characteristics and realizes closed-loop optimization through real-time data flow.

[0178] Prior art method: Use traditional PID control combined with fixed threshold protection strategy, control parameters are preset based on device rated values, no adaptive adjustment function. Monitoring relies only on basic current and voltage sensors, no load characteristic identification capability.

[0179] The test period was 3 months, during which various operating conditions were simulated, including load surge, voltage fluctuation and device switching. The data acquisition system recorded the following key performance indicators.

[0180] The effect comparison of the method of the present application and the prior art is shown in Table 1 and Figure 3 - Figure 5 as follows:

[0181] Table 1: Effect comparison table

[0182]

[0183] Result analysis:

[0184] Load compatibility and identification: The prior art method cannot automatically identify load characteristics, resulting in long manual debugging time when new devices are connected, and is prone to faults due to parameter mismatch. The method of the present application achieves high-precision load identification through non-intrusive monitoring and load characteristic fingerprinting, and automatically generates optimized control parameters, greatly improving system compatibility.

[0185] Dynamic performance: In the load surge test, the prior art method showed significant overshoot and oscillation, while the method of the present application quickly stabilized the system through control logic topology reconstruction, reducing response time by 29.2% and overshoot by 66.7%.

[0186] Energy efficiency and power quality: The method of the present application optimizes power factor and suppresses harmonics by adaptively adjusting control parameters, increasing energy efficiency by 11.8% and reducing harmonic distortion rate by 62.4%, reducing energy waste and equipment thermal stress.

[0187] Reliability and maintenance: The prior art method cannot adapt to load changes due to fixed parameters, resulting in high failure rate and rapid performance degradation. The method of the present application optimizes through online learning, continuously tracks load changes, reduces failure rate by 65.6%, and significantly delays performance degradation.

[0188] Operation and maintenance efficiency: The method of the present application shortens the control debugging time from an average of 180 minutes to 5 minutes, reducing the need for manual intervention and improving system availability.

[0189] This embodiment fully demonstrates the significant advantages of the method of the present application in solving the compatibility of hidden loads, improving the adaptive ability of the system and optimizing the long-term performance, providing an innovative solution for the field of electrical automatic control.

[0190] From the above description, it can be seen that the above-mentioned embodiments of the present application achieve the following technical effects:

[0191] 1. Achieved a fundamental change from "passive response" to "active adaptation" of the electrical control system, solving the long-standing problem of hidden load compatibility;

[0192] The load characteristic fingerprint constructed by the present application through non-invasive monitoring can accurately identify the unique electrical characteristics of various loads, and then convert the characteristic information into specific control adaptation parameters through a load characteristic deep analysis model, and finally realize the deep matching of the control system and the load characteristics through control logic topology reconstruction. A complete closed loop from load characteristic perception to control strategy generation is established, enabling the control system to have the intelligent ability of "recognizing the load, understanding the load, and adapting to the load". Actual tests show that the compatibility debugging time of the system for newly connected devices is shortened from several hours in traditional methods to 3-5 minutes, the load identification accuracy rate reaches 98.2%, and the matching degree of the control system and the load is improved by more than 40%.

[0193] 2. A "monitoring-analysis-control-optimization" full-link self-evolution system is established, breaking through the technical bottleneck of parameter fixation of traditional control systems;

[0194] The application introduces an online incremental learning mechanism based on actual operation efficiency, forming a complete self-optimization closed loop.

[0195] 3. A multi-level knowledge accumulation and migration learning framework is established, solving the technical problem of difficult control experience accumulation and reuse in complex industrial field environments.

[0196] The application realizes digitalization and systematic application of control knowledge through the linkage of the load characteristic fingerprint library, the control strategy library and the operation effect library. The system can safely share successful control experience among different devices through a federal learning framework, and at the same time, establish deep association between load characteristics, control strategies and operation effects through a knowledge graph unit. The learning achievements of a single device can be quickly replicated to the entire system, and new devices can directly call optimized control strategies for similar loads when connected, greatly improving the overall intelligent level of the system. Practical application shows that the knowledge migration mechanism improves the system optimization efficiency by more than 5 times, and the success rate of cross-device control strategy reuse reaches 87.5%, greatly reducing the labor cost of system maintenance and optimization.

[0197] Those skilled in the art will understand that the above discussion of any embodiment is merely exemplary and is not intended to suggest that the scope of the application is limited to these examples; under the concept of the application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes of different aspects of the application as described above. In order to be brief, they are not provided in details.

[0198] The application is intended to cover all such alternatives, modifications and variations as fall within the broad scope of the specification. Therefore, any omission, modification, equivalent replacement, improvement, etc. made in the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. An electric automatic control method based on artificial intelligence, characterized by, include: The real-time dynamic response characteristics of load devices in an electrical system are obtained through non-intrusive monitoring, and load characteristic fingerprints are constructed. The load characteristic fingerprint is input into the load characteristic deep analysis model, and the control adaptation parameter set is output. The control adaptation parameter set includes the key parameters that the control system needs to adjust and their optimization directions. Based on the control adaptation parameter set, the control logic topology of the electrical control system is reconstructed to achieve dynamic matching between the control system and the load characteristics. Based on the actual operating data of the reconstructed control system, control performance evaluation indicators are generated, and the load characteristic deep analysis model is optimized through online incremental learning according to the control performance evaluation indicators. 2.The artificial intelligence-based electric automatic control method of claim 1, wherein The method of acquiring real-time dynamic response characteristics of load devices in an electrical system through non-intrusive monitoring and constructing load characteristic fingerprints specifically includes: A pseudo-random binary sequence test signal with an amplitude not exceeding ±2% of the rated value is applied at the power input of the electrical system, and the voltage and current responses of the load equipment are collected synchronously. Based on the voltage and current responses, four types of real-time dynamic response features are extracted. The four types of real-time dynamic response features are normalized and encoded to generate unique load characteristic fingerprints. The load characteristic fingerprints adopt a three-level encoding structure, including a basic electrical parameter layer, a dynamic response feature layer, and an operating mode feature layer. 3.The AI-based electric automatic control method of claim 2, wherein The four types of real-time dynamic response features specifically include: Impedance frequency characteristics: The impedance amplitude and phase angle at 10 characteristic frequency points are measured in the frequency range of 0.1Hz to 2kHz. During the startup transient process, the current waveform of the first 100ms of the load startup instant is captured at a sampling rate of 100kHz, and the startup current peak value, rise time constant and oscillation damping ratio are extracted. Harmonic response characteristics, measuring the load's response amplitude and phase shift to the 3rd, 5th, and 7th characteristic harmonics; Power dynamic characteristics record the power factor change trajectory and response time during the step change of the load from no load to full load. 4.The AI-based electric automatic control method of claim 1, wherein The load characteristic deep analysis model adopts a five-layer hybrid neural network structure, including: The input adaptation layer receives four types of feature vectors from the load characteristic fingerprint, and performs normalization and feature enhancement through four dedicated feature preprocessing channels respectively. The feature fusion layer deeply fuses the four types of preprocessed features through a cross-attention mechanism to generate a unified multidimensional feature representation; The temporal memory layer employs a bidirectional long short-term memory network to capture the temporal evolution patterns and dynamic behavior patterns of load characteristics. The parsing inference layer parses the identity characteristics, health status, and adaptation requirements of the load through three parallel fully connected network branches. The output mapping layer maps the parsed results to specific control adaptation parameter sets, including PID parameter optimization suggestions, filter parameter configurations, protection threshold settings, and optimization direction indicators. The load characteristic deep analysis model is trained end-to-end by the correspondence between historical load characteristic data and optimal control parameters. During the training process, a multi-task learning strategy is adopted to simultaneously optimize three objective functions: load identification accuracy, health status assessment accuracy, and control parameter recommendation effect. 5.The artificial intelligence-based electric automatic control method of claim 4, wherein The load characteristic fingerprint is input into a load characteristic deep analysis model to output a control adaptation parameter set, specifically comprising: The three-level encoding structure features of the load characteristic fingerprint are split into four feature vectors, which are input into the input adaptation layer, wherein the basic electrical parameter layer features are input into the impedance characteristic channel, the dynamic response feature layer features are input into the transient response channel, the operation mode feature layer features are input into the behavior mode channel, and the historical operation data is input into the evolution trend channel; In the feature fusion layer, the correlation weights between the four types of feature vectors are calculated through a cross-attention mechanism to generate a fusion feature representation with context awareness capability; The time sequence memory layer receives the fused feature sequence, extracts the forward and backward time dependence through a bidirectional long short-term memory network, and outputs the dynamic evolution trajectory of the load characteristic, including the load aging trend, performance decay rate, and environmental adaptation capability; The three parallel branches of the analysis reasoning layer process the dynamic evolution trajectory, wherein: the identity recognition branch outputs the specific type and manufacturer features of the load device through a softmax classifier; the health assessment branch outputs the remaining life prediction and performance decay index of the load through a regression network; and the adaptation analysis branch outputs the optimization direction and adjustment amplitude of the control parameters through a deep reinforcement learning network; The output mapping layer synthesizes the analysis results of the three branches to generate a control adaptation parameter set containing the following contents: PID parameter optimization suggestion, filter parameter configuration, protection threshold setting, and optimization direction indication; The control adaptation parameter set further contains the confidence evaluation of each parameter and the recommended implementation time window. When the confidence is lower than a preset threshold, the system adopts a gradual adjustment strategy to implement parameter optimization in stages. 6.The artificial intelligence-based electric automatic control method of claim 1, wherein Based on the control adaptation parameter set, the control logic topology of the electrical control system is reconstructed to realize the dynamic matching of the control system and the load characteristic, specifically comprising: According to the PID parameter optimization suggestion in the control adaptation parameter set, the core algorithm structure of the controller is reconstructed, including: based on the adjustment amplitude of the proportional coefficient, the gain weight of the feedforward control channel is reconfigured; based on the optimization direction of the integral time constant, the anti-saturation mechanism and reset logic of the integral term are reconstructed; based on the correction amount of the differential coefficient, the differential lead control structure and filter compensation network are reconstructed; According to the filter parameter configuration, the topology connection of the signal processing link is reconstructed, including: based on the cutoff frequency setting, the cascade structure of the multistage low-pass filter is reconstructed; based on the center frequency configuration, the resonance point tracking mechanism of the adaptive notch filter is reconstructed; based on the filter order selection, the phase compensation network of the signal preprocessing channel is reconstructed; According to the protection threshold setting, the decision level of the system protection logic is reconstructed, including: based on the action value and time limit of the overcurrent protection, the action sequence and coordination mechanism of the hierarchical protection are reconstructed; based on the starting threshold of the overvoltage protection, the feedforward warning and hierarchical response logic of the voltage protection are reconstructed; based on the recovery hysteresis of the undervoltage protection, the adaptive judgment condition and time sequence control of system recovery are reconstructed. 7.The artificial intelligence-based electric automatic control method of claim 6, wherein, The process of reconstructing the control logic topology of the electrical control system adopts a modular replacement mechanism, which replaces the corresponding modules in the original control logic with the optimized functional modules through dynamic loading while keeping the basic architecture of the system unchanged, thereby realizing online reconstruction of the control system; The reconstructed control logic topology is mapped with the load characteristic fingerprint, and when the same or similar load characteristic fingerprint is monitored, the system automatically calls the corresponding optimized control logic topology, thereby realizing accurate matching of the control system and the load characteristics; The reconstruction process also includes a version management mechanism for the control logic topology, which records the control parameters, topology structure and running effect of each reconstruction, thereby forming a self-optimization closed loop of the control strategy. 8.The artificial intelligence-based electric automatic control method of claim 1, wherein, Based on the actual running data of the reconstructed control system, control efficiency evaluation indexes are generated, and the load characteristic deep analysis model is optimized through online incremental learning based on the control efficiency evaluation indexes, specifically including: Real-time running data of the reconstructed control system are collected, including electrical parameter data, control response data and load state data, and control efficiency evaluation indexes are generated through multi-dimensional efficiency evaluation algorithms, including stability improvement indexes, energy efficiency optimization indexes, response acceleration indexes and reliability enhancement indexes; The control efficiency evaluation indexes are input into the model incremental learning engine, and a sliding window mechanism is used to retain the running data in the recent preset time period, and the load characteristic deep analysis model is optimized through online incremental learning, and the online incremental learning process adopts a federated learning framework. 9.The artificial intelligence-based electric automatic control method of claim 8, wherein, The online incremental learning optimization of the load characteristic deep analysis model specifically includes the following steps: The real-time running data are cleaned, aligned and labeled to construct an incremental learning sample set; Based on the control efficiency evaluation results, the contribution weights of each feature in the load characteristic fingerprint are recalculated; The gradient descent algorithm is used to progressively adjust the key parameters of the model while keeping the original model structure unchanged, and the optimized model parameters are fed back to the load characteristic deep analysis model in real time, thereby forming a complete closed loop from control execution to model optimization; The new knowledge obtained through incremental learning is distilled and fused with the original model knowledge.

10. An electric automatic control system based on artificial intelligence, characterized by, The artificial intelligence-based electrical automatic control method according to any one of claims 1-9, comprising: a non-invasive monitoring module configured to obtain real-time dynamic response characteristics of load devices in an electrical system and construct a load characteristic fingerprint; a load characteristic analysis module configured to input the load characteristic fingerprint into a load characteristic deep analysis model and output a control adaptation parameter set; a control logic reconstruction module configured to reconstruct the control logic topology of the electrical control system based on the control adaptation parameter set; an online learning optimization module configured to generate control efficiency evaluation indexes based on the actual running data of the reconstructed control system and optimize the load characteristic deep analysis model through online incremental learning; a data storage and management module configured to store a load characteristic fingerprint library, a control adaptation parameter mapping library and a system running history database.

Images