Intelligent power controller and control method thereof

Abstract

The invention discloses an intelligent power controller and a control method thereof, and belongs to the technical field of intelligent power control. The method specifically comprises the following steps: S1, historical data acquisition and preprocessing: acquiring historical operation data of a power system, and performing preprocessing including abnormal value elimination, data standardization and time sequence alignment; and S2, depth feature extraction: based on the preprocessed historical data, extracting time domain basic features and physical significance features to obtain first-level features, based on the first-level features, extracting association relationships between the features and dynamic trend features to obtain second-level features, and based on the second-level features, mining and extracting depth features through nonlinear mapping and a time sequence mode to obtain third-level features. Through multi-level feature extraction, in combination with preprocessing processes of abnormal value elimination, data standardization and time sequence alignment, static operation characteristics, dynamic change rules and deep implicit modes of the power system are comprehensively captured, and high-quality and high-dimension input data are provided for a control model.

Description

Technical Field

[0001] This invention relates to the field of intelligent power control technology, and in particular to an intelligent power controller and its control method. Background Technology

[0002] With the rapid development of smart grids and the large-scale integration of distributed energy sources (photovoltaics, wind power, etc.), the operating conditions of power systems are becoming increasingly complex, placing higher demands on the control accuracy, adaptability, coordination, and fault handling capabilities of power controllers. Existing technologies often employ single control algorithms (such as standalone deep reinforcement learning, model predictive control, or droop control), which are difficult to adapt to complex and ever-changing power operating conditions. For example, a single model predictive control (MPC) exhibits lag in response when dealing with intermittent fluctuations in renewable energy sources, and a single deep reinforcement learning algorithm lacks sufficient control accuracy during the initial training phase, failing to achieve optimal control under all operating conditions. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide an intelligent power controller and its control method, which can solve the problem of insufficient adaptability of existing technology algorithms.

[0004] Technical Solution: To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, an intelligent power controller and its control method, specifically including the following steps:

[0005] S1. Historical Data Acquisition and Preprocessing: Acquire historical data of power system operation and perform preprocessing including outlier removal, data standardization, and time-series alignment.

[0006] S2. Deep Feature Extraction: Based on the preprocessed historical data, extract the basic temporal features and physical meaning features to obtain the first-level features. Based on the first-level features, extract the correlation between features and dynamic trend features to obtain the second-level features. Based on the second-level features, extract the deep features through nonlinear mapping and temporal pattern mining to obtain the third-level features.

[0007] S3. Control Model Construction and Training: Construct a fusion control model of "deep neural network-long short-term memory network" and train the model through three levels of features;

[0008] S4. Control command generation: Real-time power system operation data is collected, and the same preprocessing and deep extraction methods as in steps S1 and S2 are used to obtain real-time three-level features. The real-time three-level features are input into the "deep neural network-long short-term memory network" fusion control model to output real-time control commands.

[0009] S5. Multi-network redundant communication and command issuance: The communication layer adopts a multi-network redundant design of "fifth generation mobile communication technology + long-distance radio communication technology + local controller LAN bus", which monitors the communication status in real time and switches the communication mode to issue control commands to the execution layer.

[0010] S6. Full-Scenario Fault Diagnosis and Self-Healing: Construct a convolutional neural network fault diagnosis model, take the three-level features of historical data as input, and the fault type and fault severity as output, train the convolutional neural network fault diagnosis model, input the real-time three-level features into the convolutional neural network fault diagnosis model, output the fault type and fault severity, and implement graded processing.

[0011] S7. Model Iteration and Optimization: Periodically merge real-time running data and fault data into the historical database, and use incremental training to iteratively optimize the "deep neural network-long short-term memory network" fusion control model.

[0012] Furthermore, in step S1, the historical data includes: basic electrical data, new energy parameters, environmental parameters, and fault record data; basic electrical data includes: voltage, current, frequency, power factor, and load power; new energy parameters include: photovoltaic / wind power output and energy storage state of charge; environmental parameters include: irradiance, wind speed, and temperature; fault record data includes: fault type, fault occurrence time, fault duration, and electrical parameters at the time of the fault; outlier removal: the 3σ criterion is used to identify outliers in key parameters including voltage and current, and values ​​exceeding the mean ± 3 times the standard deviation are considered outliers, and the nearest mean is used for replacement; data standardization: Min-Max standardization is performed on parameters of different dimensions, mapping them to the [0,1] interval; time alignment: basic electrical parameters, new energy parameters, and environmental parameters are time-series matched based on timestamps to ensure data integrity at the same time node.

[0013] Furthermore, step S2 specifically includes the following steps:

[0014] S21. Based on the processed historical data, extract the basic statistical features and physical meaning features in the time domain to form primary features; among them, the basic statistical features in the time domain include: the mean, variance, peak value, peak-to-peak value, kurtosis, and skewness of voltage / current; the physical meaning features include: voltage deviation rate, current distortion rate, power fluctuation coefficient, photovoltaic / wind power output fluctuation rate, and energy storage state of charge conversion rate.

[0015] S22. Based on the primary features, the correlation and dynamic change features between features are extracted through feature interaction and time-series trend analysis to form secondary features, including: feature difference / ratio calculation, trend feature extraction, and window statistical features; wherein, feature difference / ratio calculation includes: calculating the difference between similar features and the ratio of dissimilar features in the primary features; trend feature extraction includes: performing first-order and second-order differences on key features in the primary features, including voltage deviation rate and power fluctuation coefficient, to obtain the feature change rate and acceleration; window feature extraction is to use a preset sampling period as a sliding window to calculate the maximum value, minimum value, and cumulative change of each feature in the primary features within the window;

[0016] S23. Based on the secondary features, deep features are extracted through nonlinear mapping, temporal pattern mining, and feature fusion dimensionality reduction to form tertiary features. Nonlinear mapping involves using the ReLU activation function to perform nonlinear transformation on each feature in the secondary features. Temporal pattern mining uses a sliding window to capture the temporal evolution patterns of the secondary features, extracting the sequence entropy and pattern matching degree of the features within the window. Feature fusion dimensionality reduction involves principal component analysis to reduce the dimensionality of the secondary features. Specifically, this involves calculating the covariance matrix of the secondary feature matrix and solving for the eigenvalues ​​and eigenvectors of the covariance matrix. The eigenvalues ​​are sorted from largest to smallest, and principal components are selected based on a preset contribution rate threshold. The nonlinear mapping features, temporal pattern features, and principal component features are concatenated and normalized using the L2 norm to form the tertiary features.

[0017] Furthermore, in step S3, the input to the input layer of the control model consists of three-level features, with the feature dimensions determined based on the actual extraction results. The hidden layer comprises two long short-term memory (LSM) network layers and two fully connected layers. The output layer outputs control command parameters, including energy storage charging and discharging power, voltage compensation, and photovoltaic / wind power output limits. The LSM network layers use the tanh activation function, the fully connected layers use the ReLU activation function, and the output layer uses the Linear activation function. During control model training and optimization, the mean squared error loss function is used, with the goal of minimizing the deviation between the model's output control command and the actual optimal control command. The optimizer uses the Adam optimizer with a learning rate decay strategy. The three-level features are divided into a training set and a validation set in a 7:3 ratio. The training set is input into the control model, and the predicted control command is obtained through forward propagation. The loss function is calculated, and the model parameters are updated through backpropagation. Every 50 training cycles, the model performance is verified using the validation set. If the validation set loss function value does not decrease for 10 consecutive cycles, training is stopped, and the optimal model parameters are saved.

[0018] Furthermore, in step S4, real-time power system operation data is collected, and the same preprocessing and deep extraction methods as in steps S1 and S2 are used to obtain real-time level 3 features. The real-time level 3 features are input into the control model, and the model outputs control commands through forward propagation, including but not limited to: energy storage charging and discharging power commands, voltage compensation amounts, photovoltaic / wind power output limits, and smart circuit breaker opening and closing commands.

[0019] Furthermore, in step S5, the communication layer adopts a multi-network redundancy design of "fifth-generation mobile communication technology + long-distance radio communication technology + local controller area network bus" to achieve stable transmission of control commands: remote collaborative communication prioritizes the use of the fifth-generation mobile communication technology network for in-depth feature and command interaction of multi-station and wide-area collaborative control; long-distance radio communication technology serves as a backup for the fifth-generation mobile communication technology and is suitable for scenarios with poor signal coverage in remote areas; local device-to-device communication uses the local controller area network bus for real-time command transmission between the controller and actuators such as inverters, energy storage devices, and circuit breakers; real-time detection of transmission delay and packet loss rate of each network; if the transmission delay is greater than the preset transmission delay... If the value or packet loss rate exceeds the preset packet loss rate, the system automatically switches to long-distance radio communication technology. If the local controller's LAN bus fails, wireless backup communication is triggered. When multiple smart power controllers work collaboratively, based on the deep features and control commands output by each controller model, an improved alternating direction multiplier method is used to achieve global collaborative optimization: each controller shares local real-time three-level features and model output control commands through the communication layer; the global optimization objective is set as "minimizing the voltage deviation of the entire network + minimizing the total network loss". Using the three-level features of each controller as input, the global optimization problem is decomposed into local optimization sub-problems of each agent. The local optimal solution and global coordination variables are solved through alternating iteration until the convergence condition is met.

[0020] Furthermore, in step S6, the severity of the fault is classified into minor faults, general faults, and severe faults according to preset standards. During the graded processing, for minor faults: the control command parameters are adjusted through the model; for general faults: a local isolation command is issued to isolate the faulty component, and the model regenerates control commands adapted to the remaining system; for severe faults: an emergency trip command is immediately issued to isolate the faulty area, report to the monitoring center, and start the backup power supply. After the fault is cleared, the controller automatically collects the recovered data, preprocesses it and extracts three-level features before inputting it into the model, gradually adjusts the control commands, smoothly transitions to normal operation, and supplements the fault data and recovery process data into the historical database.

[0021] Furthermore, in step S7, the specific strategy for incremental training is to freeze the parameters of the long short-term memory network layer and the first fully connected layer, unfreeze the parameters of the second fully connected layer and the input layer, use the third-level features in the merged historical database as training data, train the control model using the original training configuration, evaluate the model performance using the updated validation set, and if the weighted mean squared error loss value of the validation set is reduced by no less than a preset reduction rate threshold compared to the original model, then the new model is saved and the original model is replaced; otherwise, this incremental training is abandoned and the original model is retained.

[0022] According to another aspect of the present invention, a control method for implementing an intelligent power controller as described in any one of the above claims is provided, characterized in that it includes: a data acquisition module, a data preprocessing module, a deep feature extraction module, a control model module, a multi-network redundant communication module, a fault diagnosis and self-healing module, and a model iterative optimization module;

[0023] Data acquisition module: used to collect real-time power system operation data, and at the same time acquire historical operation data;

[0024] Data preprocessing module: Used to preprocess the collected historical and real-time data, including outlier removal, data standardization, and time-series alignment;

[0025] The deep feature extraction module is used to extract basic temporal features and physical meaning features based on preprocessed historical data to obtain first-level features. Based on the first-level features, it extracts the correlation between features and dynamic trend features to obtain second-level features. Based on the second-level features, it extracts deep features through nonlinear mapping and time series pattern mining to obtain third-level features.

[0026] Control Model Module: Used to build and store the "deep neural network-long short-term memory network" fusion control model, receive real-time three-level features, and output control commands such as energy storage charging and discharging power and voltage compensation amount through forward propagation;

[0027] Multi-network redundant communication module: It adopts a multi-network redundant design of "fifth generation mobile communication technology + long-distance radio communication technology + local controller LAN bus", monitors the communication status in real time and switches the communication mode, and sends control commands to the execution layer;

[0028] Fault diagnosis and self-healing module: used to realize fault identification and graded processing in all scenarios. It has a built-in convolutional neural network fault diagnosis model that has been trained. It takes real-time three-level features as input and outputs the fault type and severity, and performs graded processing on the fault.

[0029] Model Iteration and Optimization Module: This module is used to periodically merge real-time running data and fault data into the historical database, and to iteratively optimize the "deep neural network-long short-term memory network" fusion control model using incremental training.

[0030] Beneficial Effects: Through multi-level feature extraction, combined with preprocessing steps such as outlier removal, data standardization, and time-series alignment, the static operating characteristics, dynamic change patterns, and deep implicit modes of the power system are comprehensively captured. This provides high-quality, high-dimensional input data for the control model, improving the accuracy of control commands from the source. A fusion control model of "deep neural network-long short-term memory network" is constructed. The long short-term memory network excels at capturing long-term dependencies in time-series data, while the deep neural network possesses powerful nonlinear fitting capabilities. Together, they solve the problems of response lag and insufficient initial training accuracy of single algorithms (such as standalone model predictive control or deep reinforcement learning) under complex operating conditions. This model can accurately adapt to the complex operating conditions of the power system after the integration of distributed energy resources (photovoltaics and wind power), achieving optimal control across all scenarios. It adopts a multi-network redundancy design of "fifth-generation mobile communication technology + long-distance radio communication technology + local controller LAN bus" to flexibly switch and adapt to different scenarios such as remote wide-area collaborative control, weak signal in remote areas, and real-time interaction of local devices. At the same time, it monitors the transmission latency and packet loss rate of each network in real time, and automatically switches to the backup network when the switching conditions are met to ensure uninterrupted and low-latency transmission of control commands, and to ensure the stability and continuity of the control link.

[0031] A fault diagnosis model based on convolutional neural networks, using deep three-level features as input, can accurately identify fault types and severity. Combined with a hierarchical handling strategy of "minor fault - general fault - severe fault," it achieves rapid fault location, intelligent isolation, and self-healing, minimizing the impact of faults on the power system and improving system reliability and fault tolerance. Incremental training is used for iterative model optimization, freezing mature parameters of the long short-term memory network layer and the first fully connected layer, updating only the input layer and the second fully connected layer parameters. This eliminates the need for full model retraining. After supplementing with real-time operating data and fault data, model optimization can be completed quickly, reducing computational and time costs while enabling the model to continuously adapt to changes in the power system's operating state. For scenarios involving collaborative operation of multiple intelligent power controllers, an improved alternating direction multiplier method is adopted. The global optimization objective of "minimizing the entire network voltage deviation + minimizing total network loss" is decomposed into local optimization sub-problems for each controller. Through the communication layer sharing real-time features and control commands, local optimal solutions and global coordination variables are iteratively solved alternately, achieving global optimal control of the power system across multiple distribution areas and a wide range, improving overall operating efficiency. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating the method. Detailed Implementation

[0033] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] Example 1:

[0035] Historical data acquisition and preprocessing (S1)

[0036] Data Acquisition: Obtain the historical operating data of the hybrid power supply system over the past 12 months. The basic electrical data includes three-phase voltage, three-phase current, system frequency, power factor, and load power; new energy parameters include photovoltaic array output, wind turbine generator output, and energy storage battery state of charge; environmental parameters include irradiance, wind speed, and ambient temperature; and fault record data includes 3 short-circuit faults, 5 overload faults, and 4 voltage drop faults, along with the occurrence time, duration, and electrical parameters such as voltage, current, and power at the time of each fault.

[0037] Outlier removal: The 3σ criterion is used to identify outliers in key parameters. Taking three-phase voltage as an example, its historical data mean is 10 kV and the standard deviation is 0.2 kV. Voltage data below 9.4 kV or above 10.6 kV are identified as outliers, and the average voltage value of the two adjacent moments before and after the outlier is used to replace it.

[0038] Data standardization: Min-Max standardization is performed on all parameters with different dimensions, mapping parameters such as voltage, current, and light intensity to a range of 0 to 1. For example, 10 kV voltage is standardized to 0.5, and 1200 watts per square meter of light intensity is standardized to 1.0.

[0039] Time alignment: Based on timestamps (accurate to the second), basic electrical parameters, new energy parameters, and environmental parameters are matched in time sequence to ensure that each second-level time node contains complete data such as voltage, current, photovoltaic output, and irradiance, and to eliminate missing data due to mismatched timestamps.

[0040] Deep feature extraction (S2)

[0041] First-level feature extraction (S21): Based on the preprocessed historical data, extract the basic statistical features and physical meaning features in the time domain. The basic statistical features in the time domain include the mean, variance, peak value, peak-to-peak value, kurtosis, and skewness of voltage, and the mean, variance, and peak value of current, etc.; the physical meaning features include voltage deviation rate, current distortion rate, power fluctuation coefficient, photovoltaic output fluctuation rate, and energy storage state of charge transition rate.

[0042] Secondary feature extraction (S22): Feature interaction and time-series trend analysis are performed based on primary features. Feature difference is calculated as the difference between voltage deviation rates at adjacent time points, and feature ratio is calculated as the ratio of voltage deviation rate to power fluctuation coefficient; trend feature extraction is performed by performing first-order difference on voltage deviation rate to obtain the rate of change, and second-order difference on power fluctuation coefficient to obtain acceleration; window statistical features are set with a sampling period of 60 seconds as a sliding window, and the maximum, minimum, and cumulative change of photovoltaic power output fluctuation rate within the window are calculated.

[0043] Level 3 Feature Extraction (S23): Deep processing based on Level 2 features. A linear rectified activation function is used to perform nonlinear transformation on all features in Level 2; a 60-second sliding window is used to capture the temporal evolution pattern of Level 2 features, extracting the sequence entropy and pattern matching degree of features within the window; Principal component analysis is performed to reduce the dimensionality of the Level 2 features, calculating the covariance matrix of the Level 2 feature matrix, solving for eigenvalues ​​and eigenvectors, and selecting the top 12 principal components with a cumulative contribution rate of 85%; the nonlinear mapping features, temporal pattern features, and principal component features are concatenated and normalized using L2 to form Level 3 features.

[0044] Control model construction and training (S3)

[0045] Model Construction: A fusion control model of "deep neural network-long short-term memory network" is constructed. The input layer receives three levels of features; the hidden layer contains two long short-term memory network layers (128 neurons per layer) and two fully connected layers (64 neurons per layer); the output layer outputs control command parameters, including energy storage charging and discharging power, voltage compensation amount, and photovoltaic / wind power output limit amount.

[0046] Activation function and optimization configuration: The long short-term memory network layer uses the hyperbolic tangent activation function, the fully connected layer uses the linear rectified activation function, and the output layer uses the linear activation function; the loss function uses the mean squared error loss function, the optimizer uses the adaptive momentum estimation optimizer, the initial learning rate is 0.001, and the learning rate decay strategy is adopted (decaying to 0.9 of the original value every 100 training epochs).

[0047] Model training: The three-level features are divided into a training set (features corresponding to 70% of historical data) and a validation set (features corresponding to 30% of historical data) in a 7:3 ratio. The training set is input into the model for forward propagation to obtain the predicted control command. The mean squared error loss value between the predicted value and the actual optimal control command is calculated, and the model parameters are updated through backpropagation. The model performance is verified using the validation set every 50 training cycles. When the training reaches 320 cycles, if the validation set loss function value does not decrease for 10 consecutive cycles (stabilizing at 0.002), training is stopped, and the optimal model parameters are saved.

[0048] Control command generation (S4)

[0049] Real-time system operation data is collected, including the current three-phase voltage of 10.1 kV, current of 900 amps, photovoltaic output of 8 MW, wind power output of 3 MW, energy storage state of charge of 60%, irradiance of 900 W / m², and wind speed of 6 m / s. The same preprocessing method as in step S1 (outlier removal, Min-Max standardization, and time alignment) and the same three-level feature extraction method as in step S2 are used to obtain real-time three-level features. The real-time three-level features are input into the trained fusion control model, and control commands are output through forward propagation: energy storage charging power of 8 MW, voltage compensation of -0.1 kV, photovoltaic output limit of 95%, wind power output limit of 100%, and the smart circuit breaker remains closed.

[0050] Redundant communication and command issuance across multiple networks (S5)

[0051] The communication layer adopts a multi-network redundancy design of "5G mobile communication technology + long-distance radio communication technology + local controller area network bus". Remote collaborative communication prioritizes 5G mobile communication technology for deep feature and control command interaction between the controller and the park's power monitoring center; long-distance radio communication technology serves as a backup for 5G mobile communication technology, adapting to scenarios with weak signal coverage in the park's edge areas; local device communication uses the controller area network bus to achieve real-time command transmission between the controller and actuators such as photovoltaic inverters, energy storage battery packs, and smart circuit breakers. Real-time monitoring of network status is implemented: the preset transmission delay for 5G mobile communication technology is 50 milliseconds, and the preset packet loss rate is 1%. When the 5G mobile communication technology transmission delay reaches 65 milliseconds, it automatically switches to long-distance radio communication technology; if the local controller area network bus fails (e.g., communication interruption), wireless backup communication (long-distance radio communication technology) is immediately triggered. Since this embodiment is a single-controller control scenario, multi-controller collaborative optimization is not currently involved.

[0052] Full-scenario fault diagnosis and self-healing (S6)

[0053] The convolutional neural network fault diagnosis model has been trained using three-level features from historical fault data. The input is real-time three-level features, and the output is the fault type and severity. When a sudden voltage drop to 9.2 kV and a sudden current surge to 1400 A are detected in real-time, the real-time three-level features are input into the fault diagnosis model, outputting the fault type as "line overload fault" and the severity as "general fault." Following a tiered processing strategy, a local isolation command is issued, controlling the intelligent circuit breaker corresponding to the overloaded line to trip via the controller's local area network bus, isolating the faulty component. Simultaneously, the model re-receives real-time operating data from the remaining system, extracts the three-level features, and generates adaptation commands: adjusting the energy storage discharge power to 10 MW, voltage compensation to 0.3 kV, and limiting photovoltaic output to 80%, ensuring the stable operation of the remaining system. After the fault is resolved (overload problem is solved), the controller automatically collects the restored system data (voltage restored to 10.0 kV, current restored to 850 amps), inputs it into the model after preprocessing and three-level feature extraction, gradually adjusts the control commands to the normal state, and adds the occurrence time, duration, fault parameters, and recovery process data of this overload fault to the historical database.

[0054] Model Iteration and Optimization (S7)

[0055] At the end of each month, the real-time operational data (including normal operation data and fault data) of the month is merged into the historical database, and the fusion control model is iteratively optimized using incremental training. The incremental training strategy is as follows: freeze the parameters of the long short-term memory network layer and the first fully connected layer (which have been trained and are mature), and unfreeze the parameters of the input layer and the second fully connected layer; use the third-level features in the merged historical database as training data, and train using the original training configuration (mean squared error loss function, adaptive momentum estimation optimizer, and learning rate decay strategy); evaluate the model performance using the updated validation set (30% of the new data). If the weighted mean squared error loss value of the validation set decreases from 0.002 in the original model to 0.0016 (a reduction rate of 20%, which is higher than the preset reduction rate threshold of 10%), then the new model is saved and the original model is replaced; if the reduction rate does not reach the threshold, then this incremental training is abandoned and the original model is retained.

[0056] Example 2:

[0057] Historical data acquisition and preprocessing (S1)

[0058] Data Acquisition: Obtain historical operating data for the past 18 months for the three transformer substations. Basic electrical data includes line voltage, phase current, system frequency, power factor, and load power for each substation. New energy parameters include the output of photovoltaic power stations and the state of charge of energy storage systems in each substation. Environmental parameters include the irradiance, wind speed, and ambient temperature for each substation. Fault record data includes 4 single-phase grounding faults, 6 transformer overload faults, and 2 line short-circuit faults in each substation. Record the type, occurrence time, duration, and electrical parameters at the time of each fault.

[0059] Outlier removal: The 3σ criterion is used to identify outliers. Taking the line voltage of transformer area A as an example, the mean is 10.1 kV and the standard deviation is 0.25 kV. Voltage data below 9.35 kV or above 10.85 kV are identified as outliers and replaced with the mean voltage at the nearest time.

[0060] Data standardization: Perform Min-Max standardization on all parameters, mapping them to the 0 to 1 range. For example, standardize 10 kV voltage to 0.48 and standardize 1100 W / m² light intensity to 1.0.

[0061] Time sequence alignment: Based on second-level timestamps, the basic electrical parameters, new energy parameters, and environmental parameters of the three transformer substations are matched in a unified time sequence to ensure that all relevant data of the three transformer substations correspond completely under the same second-level time node, and missing data with inconsistent timestamps are deleted.

[0062] Deep feature extraction (S2)

[0063] First-level feature extraction (S21): Extract the basic time-domain statistical features and physical significance features of each transformer area. The basic time-domain statistical features include the mean, variance, peak value, peak-to-peak value, kurtosis, and skewness of the voltage of transformer area B, and the mean, variance, and peak value of the current, etc.; the physical significance features include the voltage deviation rate, current distortion rate, power fluctuation coefficient, photovoltaic output fluctuation rate, and energy storage state of charge conversion rate of each transformer area.

[0064] Secondary feature extraction (S22): Feature interaction and time-series trend analysis are performed based on primary features. Feature difference is calculated as the difference in voltage deviation rate between transformer area A and transformer area B at the same time. Feature ratio is calculated as the ratio of energy storage state of charge conversion rate to photovoltaic power output fluctuation rate. Trend feature extraction is performed by performing first-order difference on the power fluctuation coefficient to obtain the rate of change, and second-order difference on the voltage deviation rate to obtain the acceleration. Window statistical features are set with a sliding window of 90 seconds to calculate the maximum, minimum, and cumulative change of current distortion rate for each transformer area within the window.

[0065] Third-level feature extraction (S23): The second-level features are nonlinearly transformed using a linear rectified activation function, converting all negative features to 0 while keeping positive features unchanged; a 90-second sliding window is used to capture the temporal evolution pattern of the second-level features, and sequence entropy and pattern matching degree are extracted; principal component analysis is performed on the second-level features to reduce dimensionality, and the top 15 principal components with a cumulative contribution rate of 88% are selected; the nonlinear mapping features, temporal pattern features and principal component features are concatenated and normalized by the L2 norm to form the third-level features.

[0066] Control model construction and training (S3)

[0067] Model Construction: The intelligent power controller of each transformer area constructs a "deep neural network-long short-term memory network" fusion control model. The input layer receives three levels of features; the hidden layer contains two long short-term memory network layers (156 neurons per layer) and two fully connected layers (80 neurons per layer); the output layer outputs control command parameters, including energy storage charging and discharging power, voltage compensation amount, and photovoltaic output limit amount.

[0068] Activation function and optimization configuration: The long short-term memory network layer uses the hyperbolic tangent activation function, the fully connected layer uses the linear rectified activation function, and the output layer uses the linear activation function; the loss function is the mean squared error loss function, the optimizer is the adaptive momentum estimation optimizer, the initial learning rate is 0.0015, and the learning rate decay strategy is to decay to 0.85 of the original rate every 150 training epochs.

[0069] Model training: The three-level features of each transformer area are divided into training and validation sets in a 7:3 ratio and input into the model for training. Forward propagation is used to obtain the predicted control commands, and the parameters are updated by backpropagation after calculating the loss value. The validation set is used to verify the model every 50 training cycles. When the training reaches 400 cycles, the validation set loss value is stable at 0.0025 for 10 consecutive cycles. Training is stopped, and the optimal model parameters for each transformer area are saved.

[0070] Control command generation (S4)

[0071] Real-time operational data from three transformer substations are collected, including voltage, current, photovoltaic output, and energy storage state of charge for substation A; voltage, current, photovoltaic output, and energy storage state of charge for substation B; and voltage, current, photovoltaic output, and energy storage state of charge for substation C. After the same preprocessing as in step S1 and the same three-level feature extraction as in step S2, real-time three-level features for the three substations are obtained. The real-time three-level features of each substation are input into the corresponding fusion control model, and initial control commands are output: Substation A: energy storage charging power 3 MW, voltage compensation -0.2 kV, photovoltaic output limit 90%; Substation B: energy storage discharging power 4 MW, voltage compensation 0.3 kV, photovoltaic output limit 100%; Substation C: energy storage charging power 2.5 MW, voltage compensation -0.3 kV, photovoltaic output limit 95%.

[0072] Redundant communication and command issuance across multiple networks (S5)

[0073] The communication layer adopts a multi-network redundancy design of "5G mobile communication technology + long-distance radio communication technology + local controller area network bus". Remote collaborative communication between the three distribution area controllers and the county power dispatch center prioritizes 5G mobile communication technology to transmit real-time level-three characteristics and initial control commands for each distribution area; long-distance radio communication technology serves as a backup, adapting to the signal coverage scenario of the suburban distribution area (Distribution Area C); commands are transmitted between local devices (controllers and photovoltaic inverters, energy storage systems, and distribution transformers) using the controller area network bus. Real-time network status monitoring: the preset transmission delay for 5G mobile communication technology is 60 milliseconds, and the preset packet loss rate is 1.5%. When the 5G mobile communication technology transmission delay for Distribution Area C is detected to be 72 milliseconds, the system switches to long-distance radio communication technology; if the controller area network bus for Distribution Area A fails, the long-distance radio communication technology backup is triggered. Since this is a multi-controller collaborative scenario, each controller shares local real-time level-three characteristics and initial control commands through the communication layer. An improved alternating direction multiplier method is adopted to achieve global collaborative optimization: the global optimization objective is set as "minimizing the overall network voltage deviation + minimizing the total network loss". The global optimization problem is decomposed into local optimization sub-problems for three transformer areas. The local optimal solution and global coordination variables are solved through alternating iterations. After 10 iterations, the convergence condition is met, and the final control commands are output: transformer area A: energy storage charging power 3.5 MW, voltage compensation -0.15 kV, photovoltaic output limit 92%; transformer area B: energy storage discharging power 4.5 MW, voltage compensation 0.25 kV, photovoltaic output limit 100%; transformer area C: energy storage charging power 2 MW, voltage compensation -0.25 kV, photovoltaic output limit 98%. These commands are then sent to the execution layer through the corresponding communication network.

[0074] Full-scenario fault diagnosis and self-healing (S6)

[0075] When an overload fault occurs in the distribution transformer of area B, the real-time three-level feature input to the convolutional neural network fault diagnosis model outputs the fault type as "transformer overload fault" and the severity as "severe fault". Following a tiered handling strategy, an emergency trip command is immediately issued, controlling the smart circuit breaker corresponding to the transformer in area B to trip and isolate the faulty area. Simultaneously, the county power dispatch center is notified, and the backup diesel generator in area B is activated to supply power (ensuring power to critical loads). After the fault is resolved (by replacing the overloaded component), the controller automatically collects the operational data of area B after recovery (voltage 10.0 kV, current 720 A, photovoltaic output 3.8 MW), preprocesses and extracts the three-level features, inputs it into the model, gradually adjusts the control commands to a normal state, and supplements the historical database with the fault data and recovery process data.

[0076] Model Iteration and Optimization (S7)

[0077] Every two months, the real-time operation data and fault data of the three transformer substations are merged into their respective historical databases. Incremental training is used to optimize the models: the parameters of the long short-term memory network layer and the first fully connected layer of each model are frozen, and the parameters of the input layer and the second fully connected layer are unfrozen; the merged three-level features are used as training data, and training is performed according to the original training configuration; the performance is evaluated using the updated validation set. If the weighted mean squared error loss value of the model in transformer substation A decreases from 0.0025 to 0.002 (a reduction rate of 20%, which is higher than the preset threshold of 12%), and the reduction rates of the models in transformer substations B and C both meet the threshold requirements, then the original models of each transformer substation are saved and replaced; if the reduction rate of a model in a certain transformer substation does not meet the standard, only the original model of that transformer substation is retained, and the models of other transformer substations that meet the standard are updated.

[0078] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A control method for an intelligent power controller, characterized in that, Specifically, the following steps are included: S1. Historical Data Acquisition and Preprocessing: Acquire historical data of power system operation and perform preprocessing including outlier removal, data standardization, and time-series alignment. S2. Deep Feature Extraction: Based on the preprocessed historical data, extract the basic temporal features and physical meaning features to obtain the first-level features. Based on the first-level features, extract the correlation between features and dynamic trend features to obtain the second-level features. Based on the second-level features, extract the deep features through nonlinear mapping and temporal pattern mining to obtain the third-level features. S3. Control Model Construction and Training: Construct a "deep neural network-long short-term memory network" fusion control model and train the model using three levels of features; S4. Control command generation: Real-time power system operation data is collected, and the same preprocessing and deep extraction methods as in steps S1 and S2 are used to obtain real-time three-level features. The real-time three-level features are input into the "deep neural network-long short-term memory network" fusion control model to output real-time control commands. S5. Multi-network redundant communication and command issuance: The communication layer adopts a multi-network redundant design of "fifth generation mobile communication technology + long-distance radio communication technology + local controller LAN bus", which monitors the communication status in real time and switches the communication mode to issue control commands to the execution layer. S6. Full-Scenario Fault Diagnosis and Self-Healing: Construct a convolutional neural network fault diagnosis model, take the three-level features of historical data as input, and the fault type and fault severity as output, train the convolutional neural network fault diagnosis model, input the real-time three-level features into the convolutional neural network fault diagnosis model, output the fault type and fault severity, and implement graded processing. S7. Model Iteration and Optimization: Periodically merge real-time operating data and fault data into the historical database, and use incremental training to iteratively optimize the "deep neural network-long short-term memory network" fusion control model.

2. The control method for an intelligent power controller according to claim 1, characterized in that: In step S1, historical data includes: basic electrical data, new energy parameters, environmental parameters, and fault record data; basic electrical data includes: voltage, current, frequency, power factor, and load power; new energy parameters include: photovoltaic / wind power output and energy storage state of charge; environmental parameters include: irradiance, wind speed, and temperature; fault record data includes: fault type, fault occurrence time, fault duration, and electrical parameters at the time of the fault; outlier removal: the 3σ criterion is used to identify outliers in key parameters including voltage and current, and values ​​exceeding the mean ± 3 times the standard deviation are considered outliers, and the nearest mean is used for replacement; data standardization: Min-Max standardization is performed on parameters of different dimensions, mapping them to the [0,1] interval; time alignment: basic electrical parameters, new energy parameters, and environmental parameters are time-series matched based on timestamps to ensure data integrity at the same time node.

3. The control method for an intelligent power controller according to claim 1, characterized in that: Step S2 specifically includes the following steps: S21. Based on the processed historical data, extract the basic statistical features and physical meaning features in the time domain to form primary features; among them, the basic statistical features in the time domain include: the mean, variance, peak value, peak-to-peak value, kurtosis, and skewness of voltage / current; the physical meaning features include: voltage deviation rate, current distortion rate, power fluctuation coefficient, photovoltaic / wind power output fluctuation rate, and energy storage state of charge conversion rate. S22. Based on the primary features, the correlation and dynamic change features between features are extracted through feature interaction and time-series trend analysis to form secondary features, including: feature difference / ratio calculation, trend feature extraction, and window statistical features; wherein, feature difference / ratio calculation includes: calculating the difference between similar features and the ratio of dissimilar features in the primary features; trend feature extraction includes: performing first-order and second-order differences on key features in the primary features, including voltage deviation rate and power fluctuation coefficient, to obtain the feature change rate and acceleration; window feature extraction is to use a preset sampling period as a sliding window to calculate the maximum value, minimum value, and cumulative change of each feature in the primary features within the window; S23. Based on the secondary features, deep features are extracted through nonlinear mapping, temporal pattern mining, and feature fusion dimensionality reduction to form tertiary features. Nonlinear mapping involves using the ReLU activation function to perform nonlinear transformation on each feature in the secondary features. Temporal pattern mining uses a sliding window to capture the temporal evolution patterns of the secondary features, extracting the sequence entropy and pattern matching degree of the features within the window. Feature fusion dimensionality reduction involves principal component analysis to reduce the dimensionality of the secondary features. Specifically, this involves calculating the covariance matrix of the secondary feature matrix and solving for the eigenvalues ​​and eigenvectors of the covariance matrix. The eigenvalues ​​are sorted from largest to smallest, and principal components are selected based on a preset contribution rate threshold. The nonlinear mapping features, temporal pattern features, and principal component features are concatenated and normalized using the L2 norm to form the tertiary features.

4. The control method for an intelligent power controller according to claim 1, characterized in that: In step S3, the input to the input layer of the control model consists of three-level features, with the feature dimensions determined based on the actual extraction results. The hidden layer comprises two long short-term memory (LSM) network layers and two fully connected layers. The output layer outputs control command parameters, including energy storage charging and discharging power, voltage compensation, and photovoltaic / wind power output limits. The LSM network layers use the tanh activation function, the fully connected layers use the ReLU activation function, and the output layer uses the Linear activation function. During control model training and optimization, the mean squared error loss function is used, with the goal of minimizing the deviation between the model's output control command and the actual optimal control command. The optimizer uses the Adam optimizer with a learning rate decay strategy. The three-level features are divided into a training set and a validation set in a 7:3 ratio. The training set is input into the control model, and the predicted control command is obtained through forward propagation. The loss function is calculated, and the model parameters are updated through backpropagation. Every 50 training cycles, the model performance is verified using the validation set. If the validation set loss function value does not decrease for 10 consecutive cycles, training is stopped, and the optimal model parameters are saved.

5. The control method for an intelligent power controller according to claim 1, characterized in that: In step S4, real-time power system operation data is collected, and the same preprocessing and deep extraction methods as in steps S1 and S2 are used to obtain real-time level 3 features. The real-time level 3 features are input into the control model, and the model outputs control commands through forward propagation, including but not limited to: energy storage charging and discharging power commands, voltage compensation amounts, photovoltaic / wind power output limits, and smart circuit breaker opening and closing commands.

6. The control method for an intelligent power controller according to claim 1, characterized in that: In step S5, the communication layer adopts a multi-network redundancy design of "fifth-generation mobile communication technology + long-distance radio communication technology + local controller area network bus" to achieve stable transmission of control commands: remote collaborative communication prioritizes the fifth-generation mobile communication technology network for in-depth feature and command interaction of multi-station and wide-area collaborative control; long-distance radio communication technology serves as a backup for fifth-generation mobile communication technology and is suitable for scenarios with poor signal coverage in remote areas; local device-to-device communication uses the local controller area network bus for real-time command transmission between the controller and actuators such as inverters, energy storage devices, and circuit breakers; real-time detection of transmission delay and packet loss rate of each network; if the transmission delay exceeds the preset value or If the packet loss rate exceeds the preset value, the system automatically switches to long-distance radio communication technology. If the local controller's LAN bus fails, wireless backup communication is triggered. When multiple smart power controllers work collaboratively, a modified alternating direction multiplier method is used to achieve global collaborative optimization based on the deep features and control commands output by each controller model. Each controller shares local real-time three-level features and model-output control commands through the communication layer. The global optimization objective is set as "minimizing the voltage deviation of the entire network + minimizing the total network loss". Using the three-level features of each controller as input, the global optimization problem is decomposed into local optimization sub-problems of each agent. The local optimal solution and global coordination variables are solved through alternating iterations until the convergence condition is met.

7. The control method for an intelligent power controller according to claim 1, characterized in that: In step S6, the severity of the fault is classified into minor fault, general fault, and severe fault according to a preset standard; during the graded processing, for minor faults: the control command parameters are adjusted by the model. General faults: Issue a local isolation command to isolate the faulty component, and the model regenerates control commands adapted to the remaining system; Severe faults: Immediately issue an emergency trip command to isolate the faulty area, report to the monitoring center, and start the backup power supply; After the fault is cleared, the controller automatically collects the recovered data, inputs it into the model after preprocessing and three-level feature extraction, gradually adjusts the control commands, smoothly transitions to normal operation, and supplements the fault data and recovery process data into the historical database.

8. The control method for an intelligent power controller according to claim 5, characterized in that: In step S7, the specific strategy for incremental training is to freeze the parameters of the long short-term memory network layer and the first fully connected layer, unfreeze the parameters of the second fully connected layer and the input layer, use the third-level features in the merged historical database as training data, train the control model using the original training configuration, evaluate the model performance using the updated validation set, and if the weighted mean squared error loss value of the validation set is reduced by no less than the preset reduction rate threshold compared with the original model, then the new model is saved and the original model is replaced; otherwise, this incremental training is abandoned and the original model is retained.

9. An intelligent power controller, characterized in that, The controller is used to implement the control method of an intelligent power controller as described in any one of claims 1-8, comprising: a data acquisition module, a data preprocessing module, a deep feature extraction module, a control model module, a multi-network redundant communication module, a fault diagnosis and self-healing module, and a model iterative optimization module; Data acquisition module: used to collect real-time power system operation data, and at the same time acquire historical operation data; Data preprocessing module: Used to preprocess the collected historical and real-time data, including outlier removal, data standardization, and time-series alignment; The deep feature extraction module is used to extract basic temporal features and physical meaning features based on preprocessed historical data to obtain first-level features. Based on the first-level features, it extracts the correlation between features and dynamic trend features to obtain second-level features. Based on the second-level features, it extracts deep features through nonlinear mapping and time series pattern mining to obtain third-level features. Control Model Module: Used to build and store the "deep neural network-long short-term memory network" fusion control model, receive real-time three-level features, and output control commands such as energy storage charging and discharging power and voltage compensation amount through forward propagation; Multi-network redundant communication module: It adopts a multi-network redundant design of "fifth generation mobile communication technology + long-distance radio communication technology + local controller LAN bus", monitors the communication status in real time and switches the communication mode, and sends control commands to the execution layer; Fault diagnosis and self-healing module: used to realize fault identification and graded processing in all scenarios. It has a built-in convolutional neural network fault diagnosis model that has been trained. It takes real-time three-level features as input and outputs the fault type and severity, and performs graded processing on the fault. Model Iteration and Optimization Module: This module is used to periodically merge real-time running data and fault data into the historical database, and to iteratively optimize the "deep neural network-long short-term memory network" fusion control model using incremental training.

Images