Deep learning based active power distribution network real-time voltage control method

Abstract

The application discloses a deep learning-based active power distribution network real-time voltage control method, relates to the technical field of power grid intelligent control, and comprises the following steps: collecting multi-source power grid operation data in real time, wherein the power grid operation data refers to bus voltage amplitude and circuit breaker switch state; performing characteristic density wave analysis on the bus voltage amplitude to output a pulse time sequence; extracting voltage deviation characteristics based on actuator feedback data; constructing a pulse neural network, combining a dynamic Hebb learning rule to update network weights; generating a control increment matrix and performing a consonance region safety constraint check; processing the checked control increment matrix through a time window moving average algorithm; and outputting a smooth control instruction to an execution terminal. Through the dynamic calculation mechanism of fusing characteristic density wave analysis and circuit breaker state Shannon entropy, the application realizes accurate perception of power grid topology changes, significantly enhances the overall performance response of voltage control, shortens the response time, and improves the stability of control instructions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power grid intelligent control, in particular to a real-time voltage control method for active distribution network based on deep learning. BACKGROUND

[0002] With the continuous increase of renewable energy penetration and the large-scale grid connection of distributed energy, the operation of active distribution network presents high volatility and strong uncertainty characteristics. Real-time voltage control of active distribution network has become a key link to ensure the safe and stable operation of power system, and has core significance to support efficient consumption of new energy and improve the resilience of power grid. The dynamic control effect of bus voltage in active distribution network directly affects the power supply quality and equipment safety, and is an important technical path to optimize power grid energy efficiency and avoid voltage instability accidents. Therefore, developing real-time voltage control technology with high adaptive ability has become the research focus in the field of smart distribution network.

[0003] Traditional voltage control methods mainly rely on steady-state model optimization or PID controllers based on fixed parameters. In recent years, deep learning technology has been gradually introduced to build predictive control models to cope with new energy volatility. However, when facing dynamic scenarios such as circuit breaker state jump and feeder topology transient reconstruction, especially in complex conditions of high-penetration distributed energy areas, these methods cannot real-time integrate topology dynamic information and accurately construct safety boundaries, resulting in prominent control response lag and instruction mismatch problems, which seriously restricts control accuracy and stability. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a real-time voltage control method for active distribution network based on deep learning to solve the problems of control instruction delay, weak adaptive ability and insufficient control accuracy of existing active distribution network voltage control under high-penetration distributed energy.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] The present application provides a real-time voltage control method for active distribution network based on deep learning, which includes real-time acquisition of multi-source power grid operation data, the power grid operation data refers to bus voltage amplitude and circuit breaker switch state, characteristic density wave analysis of bus voltage amplitude to output pulse time sequence, calculation of Shannon entropy value based on circuit breaker state change, and generation of topology decision mark using spatial gradient.

[0008] The pulse time sequence is matched with the topological decision mark using time stamp interpolation and sliding time window to generate a four-dimensional dynamic correlation tensor, and an initial control vector is generated through a deep spatio-temporal convolution-circulation network, when a topological decision mark mutation intensity exceeding a preset activation threshold and a state jump is detected, a meta-learning channel is triggered;

[0009] A pre-trained meta-model is loaded and the four-dimensional dynamic correlation tensor is input, a topological correlation matrix is constructed combining the circuit breaker state, the initial control vector is optimized through multi-objective optimization, and a safe control vector is output to the actuator after safety boundary verification;

[0010] Voltage deviation features are extracted based on actuator feedback data, a pulse neural network is constructed to update network weights combined with dynamic Hebb learning rules, a control increment matrix is generated and subjected to a consonant region safety constraint verification, the verified control increment matrix is processed through a time window sliding average algorithm, and a smooth control command is output to the execution terminal.

[0011] As a preferred scheme of the active power distribution network real-time voltage control method based on deep learning, wherein: the feature density wave analysis on the bus voltage amplitude to output the pulse time sequence is,

[0012] The feature density wave is analyzed on the bus voltage amplitude to generate output feature density wave amplitude and feature phase angle, and a gradient threshold method is used to identify the local maximum value point in the feature density wave amplitude as the pulse time point;

[0013] Based on the pulse time point and the corresponding feature phase angle, the pulse time point is mapped to the phase base quantity of the power grid power frequency cycle, and the local offset of the feature phase angle is superimposed to generate the feature density wave phase;

[0014] According to the pulse time point, the voltage analysis window is determined, the voltage fluctuation features in the window are extracted and standardized, the standardized voltage fluctuation weight is generated, and the four-dimensional pulse time sequence is output.

[0015] As a preferred scheme of the active power distribution network real-time voltage control method based on deep learning, wherein the shannon entropy value is calculated based on the change of the circuit breaker state, and the topological decision mark is generated using spatial gradient, including,

[0016] Based on the change of the circuit breaker state, the bus voltage amplitude is extracted and analyzed through Clarke transformation and Hilbert envelope analysis to output sine component, cosine component and exponential decay component;

[0017] The current circuit breaker state, the sine component, the cosine component and the exponential decay component are combined into a four-dimensional dynamic feature vector, and the shannon entropy value of the four-dimensional dynamic feature vector is calculated through phase space mapping;

[0018] The spatial gradient is used to perform spatial differential operation on the Shannon entropy value to quantify the topological complexity, generate an entropy spatial gradient, and combine the current circuit breaker state to jointly generate a topological decision marker.

[0019] As a preferred scheme of the active power distribution network real-time voltage control method based on deep learning, the four-dimensional dynamic correlation tensor is generated by matching the time stamp interpolation with the sliding time window, including,

[0020] The four-dimensional pulse time series is aligned using time stamp interpolation, and a time-intensity-feature value-decision gradient four-dimensional dynamic correlation tensor is generated by a sliding time window fusion method.

[0021] As a preferred scheme of the active power distribution network real-time voltage control method based on deep learning, when the topological decision marker mutation intensity exceeds the preset activation threshold and the state jumps, the meta-learning channel is triggered, including;

[0022] Based on the second derivative change of the topological dynamic decision marker gradient field, the cumulative integral quantity in the preset time window is calculated, and when the topological decision marker mutation intensity exceeds the preset activation threshold and the state jumps, the meta-learning channel is activated.

[0023] As a preferred scheme of the active power distribution network real-time voltage control method based on deep learning, the four-dimensional dynamic correlation tensor is generated by matching the time stamp interpolation with the sliding time window, including,

[0024] The four-dimensional dynamic correlation tensor is input into the pre-trained meta-model to extract the slice of the topological decision gradient dimension and the circuit breaker state dimension, and the pre-trained meta-model is used to calculate and generate the circuit breaker on-state correlation matrix and the state transition intensity matrix;

[0025] A phase transformation operator is applied to the state transition intensity matrix to generate a phase fusion matrix, and based on the normalized voltage fluctuation weight in the four-dimensional pulse time series, a voltage fluctuation loss matrix is constructed in real time;

[0026] The circuit breaker on-state correlation matrix and the voltage fluctuation loss matrix are calculated by Hadamard product, and the basic combination matrix is output.

[0027] The phase fusion matrix and the basic combination matrix are executed by spatial and temporal feature fusion through a phase transformation matrix combination linear transformation to generate a topological correlation matrix.

[0028] As a preferred scheme of the active power distribution network real-time voltage control method based on deep learning, the optimization of the initial control vector by multi-objective optimization includes,

[0029] A three-objective optimization function is constructed based on a topological correlation matrix, a gradient separation algorithm is used to calculate the three-objective optimization function, and an initial control vector is updated in combination with an entropy space gradient;

[0030] According to the generation of the truncation parameter based on the coherent region safety constraint, the truncated mapping is performed on the updated initial control vector, and the safety control vector is output to the actuator and written into the distributed database.

[0031] As a preferred scheme of the active power distribution network real-time voltage control method based on deep learning, the voltage deviation feature is extracted based on the actuator feedback data, the pulse neural network is constructed, the network weight is updated combined with the dynamic Hebb learning rule, the control increment matrix is generated and the coherent region safety constraint is verified, the control increment matrix after verification is processed through the time window moving average algorithm, and the smooth control instruction is output, including the following steps,

[0032] Based on the voltage waveform signal feedback of the distributed database, the pulse timestamp alignment is performed with the feature density wave, the voltage deviation is calculated in the aligned time window, the sliding window statistical method is used to perform standardization processing on the voltage deviation, and the standardized voltage deviation feature vector is output;

[0033] The feature phase angle, the standardized voltage fluctuation weight, the entropy space gradient and the standardized voltage deviation feature vector are integrated and input to the four-dimensional input layer to construct the pulse neural network;

[0034] Based on the entropy space gradient norm, the learning rate of the pulse neural network is adaptively updated to generate the control increment matrix, and the coherent region safety constraint is used to verify the control increment matrix, and the smooth control instruction is output.

[0035] The present application has the following advantages:

[0036] By fusing the feature density wave analysis and the dynamic calculation mechanism of the circuit breaker state Shannon entropy, the accurate perception of power grid topology changes is realized; with the help of the four-dimensional dynamic correlation tensor, the linkage strategy of deep space convolution-meta learning is constructed, and the high-precision control vector is adaptively generated in the feeder topology jump scene; combined with the dynamic Hebb learning rule of the pulse neural network and the coherent region safety constraint verification, the closed-loop optimization of the control increment matrix and the smooth output of the instruction are driven. The above-mentioned cooperative mechanism significantly enhances the overall performance response time delay of the voltage control, and the control stability is improved to the power safety mandatory standard, and the new energy consumption capacity is substantially broken through, which provides adaptive and strong robust core control support for high-proportion renewable energy power grid. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description only constitute some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative effort.

[0038] Fig. 1 Flowchart for the deep learning-based active power distribution network real-time voltage control method.

[0039] Fig. 2 Flowchart for dynamic topology awareness.

[0040] Fig. 3 Flowchart for security optimization control.

[0041] Fig. 4 Flowchart for security incremental control. DETAILED DESCRIPTION

[0042] In order to make the above-mentioned objects, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0043] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, but the present application can also be practiced in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the scope of the present application, therefore the present application is not limited to the specific embodiments disclosed below.

[0044] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0045] Reference Figs. 1-4 For one embodiment of the present application, the embodiment provides a deep learning-based active power distribution network real-time voltage control method, comprising the following steps:

[0046] S1: Real-time acquisition of multi-source power grid operation data, the power grid operation data refers to bus voltage amplitude, circuit breaker switch state and distributed energy dynamic quantity, characteristic density wave analysis of bus voltage amplitude is performed to output pulse time sequence, shannon entropy value is calculated based on circuit breaker state change, and topology decision mark is generated using spatial gradient.

[0047] Specifically, the following steps are included:

[0048] S1.1: Collecting bus voltage amplitude, circuit breaker switch state and distributed energy dynamic quantity in real time in the process of real-time voltage control of active power distribution network, and realizing time synchronization of multiple data through PTP protocol.

[0049] S1.2: Inputting bus voltage amplitude into characteristic density wave analysis method, extracting time-frequency characteristics through characteristic modulation layer, pre-processing voltage signal through Gaussian window function, optimizing signal-to-noise ratio through time domain windowing processing and adaptive window width adjustment, and realizing frequency decoupling and component extraction through improved Clarke transformation and Hilbert transformation on this basis;

[0050] Performing FFT spectrum analysis on the extracted components, separating dominant frequency in combination with bandpass / lowpass filter; for the separated signal, using derivative weighted enhancement transient feature and wavelet threshold denoising to further suppress interference;

[0051] Integrating each pre-processing result through multi-scale feature fusion, and keeping phase consistency through phase rotation correction, outputting characteristic density wave amplitude.

[0052] Among them, the output characteristic density wave amplitude is represented as:

[0053] ;

[0054] In the formula, γ(t) is the characteristic density wave amplitude, V(τ) is the voltage instantaneous value at time point τ, σ is the Gaussian window function, t is the current time, f c is the dominant frequency, τ is the integral time variable, and (τ-t) is the time offset.

[0055] Real-time monitoring of characteristic density wave amplitude change is realized through gradient threshold method, power frequency in bus voltage signal is extracted through adaptive phase-locked loop, and dynamic tracking is realized in combination with Kalman filter;

[0056] Collecting gradient peak value sequence of the last 10 power frequency cycles, calculating the maximum gradient peak value; applying band-stop filtering and power spectrum analysis method to evaluate normalized noise level;

[0057] The dynamic threshold is linearly synthesized by three parts: the basic term (example: 25 / system power frequency), the history term (example: 0.7 times the maximum value of cycle gradient peak value), and the noise term (example: 0.3 times the square of noise level); double boundary constraints (example: taking the larger value of the lower limit 0.2 p.u. / ms and not exceeding the upper limit 2.0 p.u. / ms) are performed on the synthesized value, and the final output is the preset dynamic threshold.

[0058] When the gradient value exceeds the preset dynamic threshold, the local maximum value point is identified as the characteristic density peak and the accurate time stamp is recorded as the pulse time point.

[0059] Wherein, the characteristic density peak and record the accurate time stamp as the pulse time point is expressed as:

[0060] ;

[0061] In the formula, P is a set of characteristic density peak time stamps, t k is a pulse time point, θ grad is a gradient threshold, dt is a time variable, is a rate of change of the characteristic density amplitude at the pulse time point t k , is a curvature of the characteristic density amplitude value at the pulse time point t k , 0 is a local extreme point of the characteristic density wave, and dy is a differential variable.

[0062] Based on the time window established between each characteristic density peak and the pulse time point, the fundamental component is compared with the reference signal in the window, an interpolation algorithm is applied to improve the accuracy, and the characteristic phase angle is output.

[0063] Wherein, the output characteristic phase angle is expressed as:

[0064] ;

[0065] In the formula, φ(t k ) is a characteristic phase angle, ΔT φ is a phase calculation window, f c is a power grid fundamental frequency, dτ is a time differential, and φ is a phase angle identifier.

[0066] The pulse time point is mapped to the absolute phase base of the power frequency cycle, and the characteristic density wave phase generated by superimposing the characteristic phase angle is added.

[0067] S1.3: Based on the breaker state, the three-phase voltage is decoupled into sinusoidal components through the Clarke transformation matrix, the sinusoidal components represent the transient voltage characteristics, and the cosine components reflect the phase change trend; combined with Hilbert envelope analysis to extract exponential decay characteristics, while real-time acquisition of the breaker state (0: open / 1: closed) and encoding into digital features, the four-dimensional dynamic feature vector is constructed by fusing these four-dimensional features.

[0068] The Shannon entropy value is calculated by phase space grid mapping, the topological complexity is quantified, the spatial differential operation is performed on the Shannon entropy value using the spatial gradient to calculate the entropy space gradient, the spatial differential operation is performed using the Sobel operator to generate the entropy space gradient modulus, the topological complexity is quantified, and the topology decision flag is output when the entropy space gradient modulus exceeds the preset activation threshold and the state jumps in combination with the current breaker state change.

[0069] Wherein, the four-dimensional dynamic feature vector constructed by fusing these four-dimensional features is expressed as:

[0070] ;

[0071] wherein, is a four-dimensional dynamic feature vector, f d is a dominant oscillation frequency, S b (t) is a circuit breaker state, k is a decay coefficient, S b (t)·sin(2πf d t) is a sine wave motion feature, S b (t)·cos(2πf d t) is a cosine wave motion feature, e -кt is an exponential decay feature, b is a circuit breaker number, and d is a dominant oscillation mode.

[0072] wherein, the quantified topological complexity is represented as:

[0073] ;

[0074] wherein, H t is a topological information entropy, N c is a number of circuit breaker state combinations, p c is a probability of occurrence of a cth circuit breaker state combination, c is a total number of circuit breaker state combinations, and log2 is a standardization calculation of information entropy.

[0075] wherein, the topological decision flag is represented as:

[0076] ;

[0077] wherein, F topo is a topological decision flag, ▽H is an entropy space gradient norm, ΔS b is a circuit breaker state change amount, η H is a gradient threshold value.

[0078] Specifically, a preset activation threshold value is set, and the number of bus nodes and the number of line sections are extracted as basic parameters; a topological coefficient is selected according to the network complexity level (for example, 0.25 for a single radial network, 0.30 for a multi-loop network structure, and 0.35 for a network containing distributed energy sources), and a network size reference value is calculated according to a logarithmic function;

[0079] The ratio of the maximum transmission power to the rated capacity of the system is collected in real time as a load state parameter, and a load coefficient is selected according to the load fluctuation characteristics (for example, 0.10 for a commercial area, 0.15 for an industrial area, and 0.20 for a new energy area) to correct the operating state;

[0080] Linearly synthesize the network scale benchmark value and the running state correction term to obtain the basic threshold value, and perform differential processing through the regional type adjustment coefficient (for example, urban benchmark x 1.0, industrial zone x 0.95, rural network x 1.15, and new energy zone x 1.25);

[0081] Perform lower limit constraint processing on the synthesized value (for example, the minimum is 0.5 bits per square kilometer), and finally output the preset threshold value. The threshold value is updated in real time when the power grid structure changes, the operating parameters are adjusted quarterly, and the calculation coefficients are optimized semi-annually to form a dynamic adaptive decision benchmark for judging the topology reconstruction opportunity.

[0082] S1.4: Based on the feature density peak and the pulse time point, a symmetric analysis window is established, the trapezoidal integral algorithm is used to calculate the voltage deviation absolute value output voltage fluctuation characteristic value in the window, and standardization processing is performed, and dynamic statistical analysis is performed in combination with the historical data buffer, the current voltage fluctuation characteristic value is divided by the standard deviation to realize standardization conversion, and the standardized voltage fluctuation weight is obtained.

[0083] Among them, the voltage analysis window is determined based on the pulse time point, and the voltage fluctuation characteristic is extracted using integral operation and expressed as:

[0084] ;

[0085] In the formula, β(t k ) is the voltage fluctuation characteristic weight, V nom is the nominal voltage, ΔT is the time variable, is the normalization coefficient.

[0086] Finally, all dimension parameters are integrated, the pulse time point corresponding to the feature density peak is taken as the time dimension, the standardized voltage fluctuation weight is taken as the intensity dimension, the feature density wave is taken as the feature dimension, and the feature density wave phase is taken as the phase dimension, and a four-dimensional pulse time sequence is constructed in IEEE floating point format.

[0087] S2: The pulse time sequence and the topology decision mark are matched using timestamp interpolation and sliding time window to generate a four-dimensional dynamic correlation tensor, and an initial control vector is generated through a deep spatio-temporal convolution-cyclic network. When the mutation strength of the topology decision mark exceeds the preset activation threshold and the state jumps, the meta-learning channel is triggered.

[0088] Specifically, the following steps are included:

[0089] S2.1: Adopt IEEE 1588 PTP precision clock protocol to perform timestamp compensation and interpolation alignment on the four-dimensional pulse time points, the standardized voltage fluctuation weight, the feature density wave amplitude, and the feature density wave phase of the pulse time sequence.

[0090] The physical coordinates (longitude and latitude) of the nodes of the power grid are mapped to the Cartesian plane coordinate system by a geographic information system (GIS) to construct the spatial distribution of the physical nodes of the node power grid and realize the accurate association of electrical distance and geographical location.

[0091] The bus voltage amplitude instantaneous value, characteristic phase angle, and circuit breaker state are captured in real time by high-frequency sampling (5 kHz). Based on a 20 ms equally spaced sliding time window (corresponding to a 50 Hz power frequency complete cycle), 10 voltage fluctuation feature points accurately aligned in time are extracted in each window to form a continuous time sequence.

[0092] The historical fault database of State Grid (containing 42 short-circuit faults, 28 new energy off-grid events, and 15 load mutation cases) and PSCAD / EMTDC high-precision electromagnetic transient simulation data (new energy penetration rate of 30% to 50%, disturbance type distribution of voltage sag 45%, frequency oscillation 30%, and topology reconstruction 25%) are extracted and input into neural network training. After 10 kHz sampling (example: IEC 61869-9) and ±1 μs time synchronization (example: IEEE 1588v2) processing, a four-dimensional dynamic correlation tensor of the spatial gradient field containing four-dimensional features of spatial distribution, time sequence, pulse sequence, and topology decision flag is constructed.

[0093] S2.2: Input the four-dimensional dynamic correlation tensor into a three-layer spatio-temporal convolution structure to realize deep feature extraction. The first layer of convolution captures the coupling relationship between the standardized voltage fluctuation weight and the feature density amplitude. The second layer uses dilation convolution (dilation=2) to analyze the time propagation characteristics of the feature phase angle offset feature. The third layer of convolution fuses the feedback signal of the topology decision gradient field.

[0094] The processed features are input into a dynamic loop structure, the effective historical state is obtained through the forget gate, the real-time gradient field is fused through the input gate to update the features, and finally the 32-dimensional standardized hidden state feature vector is output through the spatio-temporal feature compression layer combined with the standardization layer. After nonlinear reinforcement, the distributed energy dynamic quantity (power fluctuation rate, frequency response rate, etc.) is coupled to generate an initial control vector that adapts to the grid operating state and is input into the actuator buffer queue.

[0095] where the 32-dimensional standardized hidden state feature vector is output by the spatio-temporal feature compression layer as follows:

[0096] ;

[0097] where, is the 32-dimensional standardized hidden state feature vector, W c is the weight matrix, is the bias vector, W f is the feature space mapping, is the uncompressed feature vector, is the vector norm.

[0098] S2.3: Calculate the spatial variation rate accumulation within a 10 ms moving window by monitoring the dynamic changes of the second derivative of the topological decision gradient field in real time, and deploy a discrete event detector to capture the topological decision marker state. When the detection of the topological decision marker mutation intensity exceeds the preset activation threshold and state jump, activate the meta-learning channel for control optimization, otherwise directly send the initial control vector to the execution end.

[0099] where, deploying a discrete event detector means:

[0100] ;

[0101] In the formula, δ topo is the topological decision marker mutation event result, [t-ΔT d ,t] is the time interval, F topo (t) is the topological state at the historical moment, F topo (t-Δt) is the topological state at the previous time t,

[0102] S3: Load the pre-trained meta-model and input the four-dimensional dynamic correlation tensor, construct the topological correlation matrix combined with the circuit breaker state, update the initial control vector through multi-objective optimization, and output the safe control vector to the actuator after safety boundary verification.

[0103] S3.1: After the activation of the meta-learning channel, load the pre-trained meta-model stored in the edge computing node and input the four-dimensional dynamic correlation tensor to extract the topological decision gradient dimension slice and the circuit breaker state dimension slice;

[0104] Generate the circuit breaker on-state correlation matrix and state transition intensity matrix through the meta-model double-channel processing; apply a phase transformation operator to the state transition intensity matrix to generate a phase fusion matrix;

[0105] Construct an exponential decay type voltage fluctuation loss matrix based on the standardized voltage fluctuation weight; perform Hadamard product operation on the circuit breaker on-state correlation matrix and the voltage fluctuation loss matrix to obtain a basic combination matrix; fuse the phase fusion matrix and the basic combination matrix through a learnable linear transformation layer to construct a topological correlation matrix.

[0106] where, constructing a topological correlation matrix means:

[0107] M topo = Φ ⊙ (S b × V loss ) × Ψ;

[0108] In the formula, M topo is the topological correlation matrix, Sb is the breaker state matrix, V loss is the voltage fluctuation loss matrix, is the Hadamard product, is the left phase transformation matrix, and is the right phase transformation matrix.

[0109] It should be noted that the voltage fluctuation loss matrix is generated by real-time calculation of voltage fluctuation characteristics.

[0110] S3.2: Based on the topology correlation matrix, a multi-objective optimization objective of voltage stability, control cost and topology coordination is established, and a gradient separation algorithm with constraints is used to dynamically update the initial control vector; through a third-order adaptive strategy DUI voltage deviation target calculation node voltage approaches the nominal value, minimizes the actuator action amplitude and strengthens the correlation of the state matrix, and at the same time, the process of dynamically updating the control vector by the gradient separation algorithm implements triple constraints.

[0111] wherein the multi-objective optimization objective is represented as:

[0112] ;

[0113] In the formula, min is the optimization objective direction is a multi-objective optimization function, is the initial control vector, is the optimized control vector, λ1 is the voltage deviation weight factor, V i is the real-time voltage of the i-th node, λ2 is the control cost weight factor, and λ3 is the topology correlation weight factor, and N is the number of grid nodes, is the control total amount change amplitude, tr(M topo ) is the topology correlation strength, g is the voltage constraint function, V min is the lower limit of voltage, V max is the upper limit of voltage, Δu max is the maximum change rate, P k real-time output power, is the rated capacity of the energy device, is the minimum working point of the energy device, h k is the output limit,

[0114] wherein the gradient separation algorithm dynamically updating the initial control vector is represented as:

[0115] ;

[0116] In the formula, is the control vector after the k-th iteration optimization, η k is the adaptive learning rate, ▽J is the multi-objective optimization direction, and ▽g is the constraint function gradient, is the control amount change, Δu is the control amount change value, sgn is the control change direction, and μ is the constraint penalty weight, To quantify the coupling strength of power grid topology.

[0117] where the triple-constraint safeguard is expressed as:

[0118] ;

[0119] where u (k+1) is the optimized control vector of the k+1th iteration, u V (k+1) is the voltage component in the control vector, u Q (k +1) is the reactive power component in the control vector, 0.95 is the lower bound of voltage constraint (unit: p.u.), 1.05 is the upper bound of voltage constraint (unit: p.u.), -Q max is the lower bound of reactive power compensation constraint, Q max is the reactive power compensation.

[0120] ;

[0121] where η k is the adaptive learning rate of the kth iteration, 0.05 is the base learning rate, e -k / 2.0 is the iteration decay factor, is the entropy gradient boost, 10.0 is the learning rate reduction trigger point, 2.0 is the learning rate boost trigger point, 0.8 is the learning rate reduction coefficient when the gradient is large, 1.05 is the learning rate boost coefficient when the gradient is small.

[0122] ;

[0123] where ▽J total is the total gradient of multi-objective optimization after injection of penalty term, 1.5 is the voltage penalty coefficient, is the partial derivative of control variable, (V-V lim ) 2 is the quadratic penalty function of voltage, V lim is the voltage clipping boundary value.

[0124] S3.3: Perform triple safety check on the optimized control vector, first enhance the stability characteristics through weight matrix conversion, then apply voltage clipping, power fluctuation, and device action rate triple protection, finally use the truncation function to constrain to the device physical safety interval, the safety control vector that passes the check is transmitted to the actuator through encrypted communication.

[0125] ;

[0126] where, is the safety control vector, is the truncation function, Ws is a security weight matrix, is a voltage control quantity lower limit, is a voltage control quantity upper limit.

[0127] It should be noted that the truncation function adopts a differential setting mechanism, and the voltage security boundary is bound to the standard hard constraint of the security value (for example, the lower limit is fixed at 0.95 p.u. and the upper limit is fixed at 0.95 p.u.).

[0128] S3.3: Perform triple security check on the optimized control vector, first convert and enhance the stability characteristics through the weight matrix, then apply voltage limiting, power fluctuation, and device action rate triple protection, and finally use the truncation function to constrain to the device physical safety interval. The security control vector that passes the check is transmitted to the actuator through encrypted communication.

[0129] S3.4: Package the security control vector using the AES-256 protocol, and transmit it synchronously through a redundant channel. Deploy a dual-channel verification mechanism in the terminal actuator, and execute the action after the control instruction is verified. At the same time, write the complete operation process (time stamp, check parameter, environment state) into the distributed database.

[0130] S4: Extract voltage deviation features based on distributed database feedback data, construct a pulse neural network combined with dynamic Hebb learning rules to update network weights, generate control increment matrix and perform consonant region safety constraint check, process the control increment matrix after check through time window sliding average algorithm, and output smooth control instruction to the execution terminal.

[0131] S4.1: Extract voltage deviation values based on distributed database storage, separate voltage deviation components through feature analysis method, then apply pulse timestamp alignment mechanism, match time reference by comparing measured voltage values, feature phase angle, and standardized voltage fluctuation weight, and use topological entropy space gradient check algorithm for triple verification of data validity, and output standardized voltage deviation feature vector.

[0132] S4.2: Construct a four-dimensional feature fusion pulse neuron input layer to integrate feature phase angle, standardized voltage fluctuation weight, entropy space gradient, and standardized voltage deviation feature.

[0133] Feature fusion is achieved through synaptic weight matrix, and in the weight update process, learning efficiency is adaptively adjusted based on Hebb learning rules according to the amplitude of entropy space gradient, and neuron firing frequency normalization method is used to suppress high-frequency oscillation, stabilizing the output control increment matrix.

[0134] wherein the adaptive adjustment of learning efficiency based on Hebb learning rules according to the amplitude of entropy space gradient is represented as:

[0135] Δwij = η · exp(-γ · ‖∇ε‖) · (v i v j - β · sgn(w ij )w 2 ij );

[0136] where Δw ij is the change of neuron weight, γ is the sensitivity of learning rate to topological change, v i v j is the neuron firing frequency correlation term, β · sgn(w ij )w 2 ij is the weight oscillation suppression term, and β is the oscillation suppression coefficient.

[0137] S4.3: Based on the control increment matrix of the neural network output, the spatial topology mapping is performed according to the impedance characteristics of the grid nodes, the matching degree of the relative value of mutual impedance and the geometric distance is calculated, the strong, medium and weak coupling regions are divided, the real-time / second-level / economic partition verification mechanism is driven, the three-dimensional safety benchmark (voltage / power / equipment) penetration risk assessment is synchronously performed, and the hierarchical execution strategy of low-risk full-quantity execution→medium-risk proportion attenuation→high-risk blocking alarm is output.

[0138] Based on the strong / medium / weak area label of spatial topology division, differential encoding is implemented: 3-bit simplified code elements are allocated to strong coupling areas, and 8-bit fault-tolerant code elements are used in weak coupling areas; the compression strength is adjusted according to the risk level: 4:1 high compression is enabled for low risk (penetration risk <1%), and the original data transmission is maintained for high risk (>5%); finally, the difference travel encoding is used to merge consecutive same values (example: [0, 0, 0, 0.1]→(0, 3), (0.1, 1)), and the depth-optimized compression instruction set is generated.

[0139] S4.4: Based on the compression instruction set, an event-driven weight allocation mechanism is established, and the control weight is dynamically weighted and fused by standardizing the voltage fluctuation weight and the entropy space gradient. The triple-time-window moving average algorithm (example: recent 3-period weight 0.7, medium-term 10-period 0.2, long-term 50-period 0.1) is used to process weight fluctuations, and the front and rear period weighted smoothing strategy (example: current instruction weight 0.6, previous period 0.3, previous period 0.1) is used to suppress mutation noise, and the smoothed control instruction is output to the execution terminal.

[0140] To sum up, the application realizes accurate perception of power grid topology changes by fusing feature density wave analysis and dynamic calculation mechanism of circuit breaker state Shannon entropy; generates high-precision control vector in feeder topology jump scene by means of linkage strategy of four-dimensional dynamic correlation tensor construction and deep space-time convolution-meta learning; and drives closed-loop optimization of control increment matrix and smooth output of instructions by combining dynamic Hebb learning rule of pulse neural network and coherent region safety constraint verification. The above-mentioned cooperative mechanism significantly enhances the overall performance of voltage control: the response time is shortened, the control stability is improved to the power safety mandatory standard, and the new energy consumption capacity is substantially broken through, thereby providing adaptive and strong-robust core control support for high-proportion renewable energy power grid.

[0141] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A deep learning based active power distribution network real-time voltage control method, characterized in that: The method comprises the steps of: Real-time acquisition of multi-source power grid operation data, wherein the power grid operation data refers to bus voltage amplitude and circuit breaker switch state, characteristic density wave analysis of bus voltage amplitude to output pulse time sequence, calculation of Shannon entropy value based on circuit breaker state change, and generation of topology decision mark using spatial gradient, Based on the change of the circuit breaker state, the bus voltage amplitude is extracted and analyzed by Clarke transformation and Hilbert envelope analysis, and the sine component, cosine component and exponential decay component are output; The current circuit breaker state, the sine component, the cosine component and the exponential decay component are combined into a four-dimensional dynamic feature vector, and the Shannon entropy value of the four-dimensional dynamic feature vector is calculated by phase space mapping; The spatial gradient is used to perform spatial differential operation on the Shannon entropy value to quantify the topology complexity, generate entropy space gradient, and jointly generate topology decision mark with the current circuit breaker state; The pulse time sequence and the topology decision mark are matched to generate a four-dimensional dynamic correlation tensor using timestamp interpolation and sliding time window, and an initial control vector is generated through deep spatio-temporal convolution-cyclic network, and when the mutation strength of the topology decision mark exceeds the preset activation threshold and the state jumps, the meta-learning channel is triggered; Load the pre-trained meta-model and input the four-dimensional dynamic correlation tensor, construct the topology correlation matrix combined with the circuit breaker state, optimize the initial control vector through multi-objective optimization, and output the safe control vector to the actuator after safety boundary verification; Based on the feedback data of the actuator, the voltage deviation features are extracted, the pulse neural network is constructed, the network weight is updated combined with the dynamic Hebb learning rule, the control increment matrix is generated and the safety constraint verification of the consonant region is performed, the control increment matrix after verification is processed through time window sliding average algorithm, and the smooth control command is output to the execution terminal.

2. The deep learning based active power distribution network real-time voltage control method of claim 1, wherein: The characteristic density wave analysis of the bus voltage amplitude to output the pulse time sequence refers to: The bus voltage amplitude is analyzed by the characteristic density wave to generate the output characteristic density wave amplitude and the characteristic phase angle, and the gradient threshold method is used to identify the local maximum value point in the characteristic density wave amplitude as the pulse time point; Based on the pulse time point and the corresponding characteristic phase angle, the pulse time point is mapped to the phase base quantity of the power grid power frequency cycle, and the local offset of the characteristic phase angle is superimposed to generate the characteristic density wave phase; According to the pulse time point, the voltage analysis window is determined, the voltage fluctuation features in the window are extracted and standardized, the standardized voltage fluctuation weight is generated, and the four-dimensional pulse time sequence is output.

3. The deep learning based active power distribution network real-time voltage control method of claim 1, wherein: The four-dimensional pulse time sequence is aligned using timestamp interpolation, and the four-dimensional dynamic correlation tensor of time-intensity-feature value-decision gradient is generated by sliding time window fusion method. When the mutation strength of the topology decision mark exceeds the preset activation threshold and the state jumps, the meta-learning channel is triggered, which comprises:

4. The deep learning based active power distribution network real-time voltage control method of claim 1, wherein: Based on the second derivative change of the topology dynamic decision mark gradient field, the cumulative integral quantity in the preset time window is calculated, and when the mutation strength of the topology decision mark in the window exceeds the preset activation threshold and the state jumps, the meta-learning channel is activated. ​ 5. The deep learning based active power distribution network real-time voltage control method of claim 1, wherein: The loading pre-training meta-model and inputting the four-dimensional dynamic correlation tensor, combining the circuit breaker state to construct the topological correlation matrix, including the following steps, The four-dimensional dynamic correlation tensor is input into the pre-training meta-model to extract the slice of the topological decision gradient dimension and the circuit breaker state dimension, and the pre-training meta-model is used to calculate and generate the circuit breaker on-state correlation matrix and the state transition intensity matrix The phase transformation operator is applied to the state transition intensity matrix to generate the phase fusion matrix, and the voltage fluctuation loss matrix is constructed in real time based on the normalized voltage fluctuation weight in the four-dimensional pulse time sequence; The Hadamard product calculation is performed on the circuit breaker on-state correlation matrix and the voltage fluctuation loss matrix, and the basic combination matrix is output. Through the phase transformation matrix combination linear transformation, the phase fusion matrix and the basic combination matrix are executed to perform space-time feature fusion to generate the topological correlation matrix.

6. The deep learning based active power distribution network real-time voltage control method of claim 1, wherein: The optimization of the initial control vector includes, Based on the topological correlation matrix, a three-objective optimization function is constructed, a gradient separation algorithm is used to calculate the three-objective optimization function, and the initial control vector is updated combined with the entropy space gradient; According to the safety constraint of the coherent region, the truncation parameter is generated, the updated initial control vector is executed to perform truncation mapping, and the safety control vector is output to the actuator and written into the distributed database.

7. The deep learning based active power distribution network real-time voltage control method of claim 1, wherein: The voltage deviation feature is extracted based on the feedback data of the actuator, the pulse neural network is constructed, the network weight is updated combined with the dynamic Hebb learning rule, the control increment matrix is generated and the coherent region safety constraint is verified, the verified control increment matrix is processed by the time window sliding average algorithm, and the smooth control instruction is output, including the following steps, Based on the voltage waveform signal feedback from the distributed database, the pulse timestamp is aligned with the feature density wave, the voltage deviation is calculated in the aligned time window, the sliding window statistical method is used to perform standardization processing on the voltage deviation, and the standardized voltage deviation feature vector is output; The feature phase angle, the normalized voltage fluctuation weight, the entropy space gradient and the standardized voltage deviation feature vector are integrated and input into the four-dimensional input layer to construct the pulse neural network; Based on the entropy space gradient norm, the learning rate of the pulse neural network is adaptively updated to generate the control increment matrix, the coherent region safety constraint is used to verify the control increment matrix, and the smooth control instruction is output.

Images