Standardized intelligent control system of transformer substation

Abstract

The invention relates to the technical field of intelligent control of transformer substations, in particular to a standardized intelligent control system of a transformer substation, which realizes real-time response to the running state of equipment by acquiring the current, voltage, power factor, temperature and humidity, operation state, insulation state and environment temperature and humidity of a power transformation unit in real time. A gradient boosting decision tree is combined with a variational auto-encoder and a self-attention mechanism, state vector features are extracted, complex abnormal states are accurately recognized, potential faults are detected in an early stage, and the monitoring efficiency and the early warning precision are greatly improved. The future power demand is accurately predicted by combining the long-short-term memory network model with historical and real-time load data, and the limitation of traditional static load management is overcome. Based on the predicted power demand, the reinforcement learning model is adopted to dynamically adjust the operation number of the power transformation units, optimal configuration of resources is achieved while the demand is met, and the overall operation efficiency of the transformer substation is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent control of substations, and in particular to a standardized intelligent control system of substations. Background Art

[0002] Substations are important nodes in the power transmission and distribution system. Their management model usually relies on manual inspections and regular maintenance, which makes it difficult to respond to changes in equipment operating status and fluctuations in load demand in a timely manner. This management method not only has problems such as low monitoring efficiency and poor equipment failure warning capabilities, but also increases operating costs.

[0003] Existing substation monitoring systems usually include basic parameter collection and fault alarm functions, but lack intelligent analysis methods for complex operating conditions in substations. Although some systems have introduced data collection and analysis modules, they are mostly based on fixed threshold alarm methods, unable to dynamically adapt to changes in equipment status, and difficult to deal with complex abnormal situations. Due to the inability to effectively identify and predict potential equipment failures, the reliability and early warning accuracy of the current system are limited. In addition, the existing systems mostly use static load distribution methods in load management, which cannot accurately predict real-time load changes and future power demand, resulting in low overall operating efficiency of the substation, making it difficult to adapt to fluctuations in power demand, and thus causing waste of resources due to too many substation units running at the same time. Summary of the invention

[0004] To solve the above problems, the present invention provides a standardized intelligent control system for substations, which solves the problems of low monitoring efficiency, poor early warning capability, static load management and waste of resources in traditional substations through intelligent monitoring, dynamic fault detection, load forecasting and reinforcement learning control strategy.

[0005] To achieve the above object, the technical solution adopted by the present invention is: A standardized intelligent control system for a substation, comprising: a distributed monitoring module, a fault detection module, a load monitoring module and an intelligent control module, wherein the distributed monitoring module and the fault detection module are connected in sequence, and the load monitoring module and the intelligent control module are connected; The distributed monitoring module is used to obtain active parameters and passive parameters of the substation unit in the substation, the active parameters include current, voltage, power factor, temperature and humidity, operating state and insulation state, and the passive parameters include ambient temperature and humidity inside and outside the substation; The fault detection module is used to receive active parameters and passive parameters of a single substation unit and standardize them to generate a substation unit state vector, perform preliminary abnormal classification based on the gradient boosting decision tree, and determine the abnormal state; according to the abnormal state, a variational autoencoder and a self-attention mechanism are used to extract features from the state vector to detect potential faults; The load monitoring module is used to obtain and store substation load data; The intelligent control module is used to predict future power demand data through historical substation load data and real-time substation load data based on a long short-term memory network, and control the operating quantity of substation units without abnormalities and potential faults through a reinforcement learning model based on the future power demand data, and the operating quantity is negatively correlated with the reward value in the reinforcement learning model.

[0006] Furthermore, the substation unit includes a main transformer, a high-voltage switch cabinet, a low-voltage switch cabinet, a circuit breaker, a busbar system, a cable connector and a lightning arrester.

[0007] Furthermore, the distributed monitoring module includes a current sensor, a voltage sensor, a temperature and humidity sensor, an insulation detector and an embedded operation monitoring system.

[0008] Furthermore, the receiving of active parameters of a single substation unit and standardizing the generation of a substation unit state vector, and performing preliminary abnormality classification based on a gradient boosting decision tree include the following steps: De-noise the active and passive parameters and fuse them to generate a high-dimensional feature matrix; Based on the high-dimensional feature matrix, an adaptive normalization model is used to generate real-time dynamic state vectors; Generate a threshold matrix based on the historical high-dimensional feature matrix set through the gradient boosting decision tree, and embed the adaptive threshold of each parameter into the state vector; The state vector is compared with the threshold matrix, and the state vector is classified based on the hierarchical path of the gradient boosting decision tree to determine the abnormal state, which includes no abnormality, fluctuation and failure.

[0009] Furthermore, the gradient boosting decision tree is constructed by the following steps: Clean and normalize the historical high-dimensional feature matrix set; Based on the parameters in the feature matrix, the key parameters that have a significant impact on anomaly classification are screened out through the SHAP value; Based on the selected key parameters, an initial decision tree model is constructed with the goal of minimizing the classification error, and the initial branch path is set according to the distribution characteristics of the parameters under different abnormal conditions; In each training iteration, the sample weights are adjusted according to the classification error of the previous round, and a new decision tree is constructed through the residual minimization strategy. The new decision tree is combined with the previous decision tree to generate a gradient boosting decision tree model.

[0010] Furthermore, the formula of the gradient boosting decision tree is as follows: ; in, is the prediction model of the mth iteration; is the prediction model of the previous round; is the learning rate; is the number of samples in the training data set; is the i-th sample; is the i-th sample The optimal weight corresponding to the residual of ; Weak decision tree generated for this round.

[0011] Furthermore, the feature extraction of the state vector using a variational autoencoder and a self-attention mechanism according to the abnormal state includes the following steps: The state vector is input into the encoding layer of the variational autoencoder, and the state vector is encoded with high-dimensional features through several layers of neural networks to generate latent variable representation; Apply the self-attention mechanism to the latent variable representation of the variational autoencoder output and calculate the attention weight of each dimension feature; The feature representation processed by the self-attention mechanism is input into the decoding layer of the variational autoencoder to generate a reconstructed state vector; The difference between the reconstructed state vector and the original state vector is calculated to determine the abnormal feature vector, and the corresponding potential fault is matched according to the abnormal feature vector.

[0012] Furthermore, the method of predicting future power demand data based on the long short-term memory network through historical substation load data and real-time substation load data includes the following steps: De-noising, smoothing and normalizing the historical load data and real-time load data, converting them into a time series data matrix, and constructing time windows in segments to generate a standardized load data input matrix; The standardized load data input matrix is ​​divided into blocks according to the time dimension, and the load amount and timestamp are mapped into a multi-dimensional feature vector sequence; Add time labels and load change rates to the feature sequences in each time window to form a time dependency matrix as the input of the long short-term memory network model; The time dependency matrix is ​​input into the long short-term memory network model, and forward and backward propagation training is performed on each input time window, and the model parameters are optimized by gradient descent; Input the real-time load data sequence into the trained long short-term memory network model, and use the memory unit to update the state of the sequence features; The future electricity demand data is output through the output layer.

[0013] Furthermore, the steps of constructing the reinforcement learning model include: The predicted future power demand data is used as the environmental input state of the reinforcement learning model to initialize the operating state and demand load of each substation unit in the substation to form the initial state matrix of the system; According to the operation status of the substation unit, a set of control actions is constructed, including increasing, decreasing or maintaining the number of substation units in operation; Construct the reward function so that the number of runs is negatively correlated with the reward value in the reinforcement learning model; By making strategic decisions based on the current state matrix and reward function in each time step, the action with the highest reward value is selected to adjust the number of substation operations; The load response, reward value, and state change data after each action are recorded and fed back to the model to update the state matrix and optimize the reinforcement learning model.

[0014] The beneficial effect of the present invention is that the present invention can timely respond to changes in the operating state of the equipment by acquiring active parameters (including current, voltage, power factor, temperature and humidity, operating state and insulation state) and passive parameters (including ambient temperature and humidity) of the substation unit in real time. The gradient boosting decision tree and variational autoencoder (variational autoencoder) are combined with the self-attention mechanism to extract features of the state vector, so as to realize accurate identification of complex abnormal states and early detection of potential faults, and effectively improve the monitoring efficiency and fault warning accuracy of the system. Different from the traditional alarm method based on fixed thresholds, the present invention adopts a machine learning algorithm, dynamically generates thresholds through an intelligent judgment module and performs abnormal classification in real time, so that the system can flexibly adapt to changes in the equipment state. By intelligently analyzing the collected multidimensional data, the system can accurately identify various operating states, support the processing of complex abnormal situations, and realize intelligent analysis and dynamic response to complex states in the substation. By introducing the long short-term memory network (LSTM) model, the historical load data and real-time load data are combined to realize accurate prediction of future power demand. The LSTM model can capture the time series characteristics of load data, ensuring accurate prediction results when power demand changes, overcoming the limitations of static load management allocation in traditional systems. Based on future power demand data, this solution uses a reinforcement learning model to control the number of substations in operation, and dynamically adjusts the number of substations in operation to ensure that power demand is met while minimizing resource waste. The reinforcement learning model continuously optimizes the control strategy by setting a reward function that is negatively correlated with the number of operations, achieving efficient resource allocation and optimal operating status, and improving the overall operating efficiency of the substation. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a structural schematic diagram of a standardized intelligent control system of a substation in the present invention.

[0016] Figure 2 It is a flow chart of the construction steps of the gradient boosting decision tree in the present invention. DETAILED DESCRIPTION

[0017] See also Figure 1-Figure 2 As shown, the present invention relates to a standardized intelligent control system for a substation, comprising: a distributed monitoring module, a fault detection module, a load monitoring module and an intelligent control module, wherein the distributed monitoring module and the fault detection module are connected in sequence, and the load monitoring module and the intelligent control module are connected; The distributed monitoring module is used to obtain active parameters and passive parameters of the substation unit in the substation, the active parameters include current, voltage, power factor, temperature and humidity, operating state and insulation state, and the passive parameters include ambient temperature and humidity inside and outside the substation; The fault detection module is used to receive active parameters and passive parameters of a single substation unit and standardize them to generate a substation unit state vector, perform preliminary abnormal classification based on the gradient boosting decision tree, and determine the abnormal state; according to the abnormal state, a variational autoencoder and a self-attention mechanism are used to extract features from the state vector to detect potential faults; The load monitoring module is used to obtain and store substation load data; The intelligent control module is used to predict future power demand data through historical substation load data and real-time substation load data based on a long short-term memory network, and control the operating quantity of substation units without abnormalities and potential faults through a reinforcement learning model based on the future power demand data, and the operating quantity is negatively correlated with the reward value in the reinforcement learning model.

[0018] In some embodiments, the standardized intelligent control system of the substation is mainly composed of a distributed monitoring module, a fault detection module, a load monitoring module and an intelligent control module. The distributed monitoring module is composed of high-precision sensors and embedded monitoring devices, and its specific function is to obtain and monitor the active and passive parameters of the substation unit in real time. To ensure the accuracy of parameter acquisition, the current sensor, voltage sensor, power factor sensor and temperature and humidity sensor are connected to the processing unit of the monitoring module through a high-precision data interface, and the analog signal is converted into a digital signal through an ADC converter internally, and further data fusion and denoising are performed. By transmitting the active parameters together with the passive parameters to the fault detection module, all-round monitoring of the equipment status is ensured. The design of the fault detection module is based on a multi-layer hardware architecture so that efficient preliminary abnormality analysis can be performed after receiving the monitoring data. First, the hardware platform of the module has built-in FPGA and GPU parallel processors to support the real-time operation of the gradient boosting decision tree algorithm. The FPGA is used to generate and update the threshold matrix in real time, and the GPU performs data classification calculations. The parameter data received by the detection module is standardized on the FPGA to generate the state vector of the substation unit, and the preliminary abnormality classification is performed through the gradient boosting decision tree under the collaborative calculation of the GPU. If an abnormal state is detected, the module will use a variational autoencoder (VAE) combined with a self-attention mechanism to extract features and filter out feature information that may be related to potential faults. The encoding layer of the variational autoencoder is deployed on the GPU to further accelerate the compression and mapping of high-dimensional features, and the self-attention mechanism is used to extract key features to identify complex fault modes, thereby improving the system's fault detection accuracy. The load monitoring module is responsible for the continuous collection and storage of load data to ensure the traceability of historical data of power loads. The module uses non-volatile storage devices (such as EEPROM and NAND flash memory). Data is synchronously written to local storage during collection and periodically transmitted to a remote database. The stored load data will be used as input to the intelligent control module for training the load forecasting model. The intelligent control module mainly realizes load forecasting and dynamic control of substations through long short-term memory networks (LSTM) and reinforcement learning models. In terms of hardware design, the prediction calculation of the LSTM model is completed through the collaboration of the CPU and GPU, and the GPU accelerates matrix operations to improve the prediction response speed. The reinforcement learning model is used to dynamically adjust the operating status of the substation unit according to the predicted load demand. Specifically, the module combines load forecast data and adjusts the number of operations through a reinforcement learning model. The reward function in the model is set to be negatively correlated, punishing too many or too few operations to ensure the system's adaptability when load demand changes. The final control signal is generated by the processing unit of the intelligent control module, and the actual number of operations of each substation is regulated by a high-precision drive circuit, thereby achieving optimal resource allocation and efficient management.

[0019] Furthermore, the substation unit includes a main transformer, a high-voltage switch cabinet, a low-voltage switch cabinet, a circuit breaker, a busbar system, a cable connector and a lightning arrester.

[0020] Specifically, the main transformer, as the core substation equipment, is connected between the incoming and outgoing lines of the substation. The current and voltage sensors in the distributed monitoring module monitor its input and output voltage, current value, and environmental parameters such as temperature and humidity. The monitoring module standardizes the operating parameters of the main transformer, generates a real-time state vector, and identifies anomalies through the gradient boosting decision tree of the fault detection module to ensure the stability and reliability of the main transformer operation.

[0021] High-voltage switchgear is used to connect or disconnect high-voltage lines. Its operating status, temperature, insulation status and other parameters are collected by sensors embedded in the equipment. In order to improve the response speed and reliability of the system, the high-voltage switchgear transmits the collected data to the fault detection module in real time through the FPGA parallel processing hardware, and extracts key abnormal features in combination with the self-attention mechanism, so as to respond quickly when a fault or abnormal operating state occurs. The low-voltage switchgear is mainly responsible for distributing electric energy to other equipment and external loads in the substation. The monitoring system arranges multi-channel temperature and humidity sensors and voltage and current sensors inside the low-voltage switchgear, and collects and stores the load status of the low-voltage switchgear in real time through the load monitoring module. The intelligent control module will dynamically adjust the working status of the low-voltage switchgear based on the load prediction results and the reinforcement learning model to ensure the efficiency of load distribution. The function of the circuit breaker is to protect the circuit and avoid equipment damage due to overload or short circuit. The circuit breaker status monitoring uses high-precision current sensors and thermistors to detect abnormal conditions such as overcurrent and temperature rise. The distributed monitoring module transmits the collected status data to the fault detection module, and classifies it through the gradient boosting decision tree. If an abnormality is found, the system immediately triggers the fault detection process and conducts in-depth analysis of potential faults. The busbar system connects various devices and is used to distribute electrical energy. The operating parameters of the busbar system, such as temperature and current carrying capacity, are obtained in real time by the distributed monitoring module and transmitted to the load monitoring module for reference during load forecasting. If the load forecast result shows that the busbar load is close to the upper limit, the intelligent control module adjusts the number of substation units in operation through the reinforcement learning model to ensure safe load distribution of the busbar. As a key component connecting various devices, the cable joint is susceptible to ambient temperature and load fluctuations, and abnormal temperature rise may occur. The system monitors the temperature changes of the cable joint through a temperature sensor, and uses the variational autoencoder model combined with the self-attention mechanism in the fault detection module to identify potential faults, providing early warning for abnormal temperature rise of the cable joint. Lightning arresters are used to protect equipment from lightning strikes. The current sensor and ground resistance monitoring module inside the arrester detect the number of discharges and ground resistance values ​​in real time. The distributed monitoring module transmits these data to the fault detection module for standardized processing to ensure that the arrester works normally when struck by lightning, and automatically analyzes and alarms when an abnormality occurs, so as to improve the system's lightning resistance and safety.

[0022] Furthermore, the distributed monitoring module includes a current sensor, a voltage sensor, a temperature and humidity sensor, an insulation detector and an embedded operation monitoring system.

[0023] Specifically, the current sensor is used to monitor the current flowing through the substation in real time to ensure that the equipment is within a safe load range. The high-precision current sensor converts the collected current signal into a digital signal through a low-noise signal amplifier and a filter circuit, and transmits it to the embedded operation monitoring system for processing. In the case of high load or abnormal current fluctuations, the current sensor will trigger a real-time alarm, prompting the fault detection module to further analyze the cause of the abnormal current. The voltage sensor is responsible for monitoring the voltage state of the substation, converting the analog voltage signal into a digital signal through a voltage divider circuit and an ADC module and then transmitting it to the monitoring system. In order to improve the accuracy of voltage monitoring, the voltage sensor uses multi-layer electrical isolation technology to avoid interference and overvoltage damage. When the voltage fluctuation exceeds the set threshold, the system will automatically trigger the fault detection module and further analyze whether there is a potential fault in combination with other parameters. The temperature and humidity sensor is used to collect temperature and humidity data inside the substation to prevent equipment aging or insulation performance degradation due to abnormal temperature and humidity. The sensor has a built-in microprocessor that can convert analog signals into standard digital signals through a protocol conversion interface and transmit them to the embedded operation monitoring system. When the temperature or humidity exceeds the standard, the system will automatically record environmental data and link the insulation detector to conduct insulation status inspection. The embedded operation monitoring system is the core part of the distributed monitoring module, which is responsible for receiving, processing and transmitting the data information and operation status data collected by each sensor. The embedded system integrates FPGA and MCU to collect various data through high-speed interfaces.

[0024] Furthermore, the receiving of active parameters of a single substation unit and standardizing the generation of a substation unit state vector, and performing preliminary abnormality classification based on a gradient boosting decision tree include the following steps: De-noise the active and passive parameters and fuse them to generate a high-dimensional feature matrix; Based on the high-dimensional feature matrix, an adaptive normalization model is used to generate real-time dynamic state vectors; Generate a threshold matrix based on the historical high-dimensional feature matrix set through the gradient boosting decision tree, and embed the adaptive threshold of each parameter into the state vector; The state vector is compared with the threshold matrix, and the state vector is classified based on the hierarchical path of the gradient boosting decision tree to determine the abnormal state, which includes no abnormality, fluctuation and failure.

[0025] In some embodiments, first, multi-level denoising is applied to active parameters and passive parameters. The system uses a wavelet transform algorithm for multi-scale denoising to separate high-frequency noise components and actual useful signals in the signal. In the initial denoising step, a low-pass filter is used to weaken the instantaneous high-frequency noise during the acquisition process to ensure the stability of the core feature data. Next, the signal is decomposed into different frequency bands through adaptive wavelet denoising, retaining low-frequency components reflecting characteristics such as current and voltage, and removing interference signals. The denoised data is fused with historical data to generate a high-dimensional feature matrix to provide an accurate feature representation. Based on the high-dimensional feature matrix, the system uses an adaptive standardization model to unify the scale of each feature. First, a statistical analysis of the historical data distribution of each parameter is performed, including dynamic updates of the mean, variance, minimum and maximum values ​​to generate a standardized feature description. The adaptive standardization model normalizes the data in the feature matrix and maps each feature to the same numerical interval, so that the data has the same scale in different feature dimensions. For example, the standardization of the current parameter increases the weight due to its large dynamic range, while the temperature parameter is given a lower weight due to its small fluctuation range. At the same time, the model dynamically adjusts the standardized weights of the parameters to cope with the impact of load fluctuations or environmental changes. For example, when the load increases, the adaptive normalization model automatically relaxes the parameter weights of current and power factor to capture more significant change trends in the state vector. Next, the system uses the trained gradient boosting decision tree (GBDT) to perform preliminary abnormal classification of the state vector. First, the GBDT model is trained based on a high-dimensional feature matrix set. Each round of iteration gradually establishes a weak classifier, and the model is iteratively optimized through the residual minimization strategy to ensure that the model can generate an accurate threshold matrix. The threshold matrix contains the upper and lower limits of each parameter in different states, representing the safety range of the system in the states of "no abnormality", "fluctuation", and "fault". For example, the thresholds of parameters such as current and insulation resistance are dynamically adjusted according to historical data and environmental characteristics. After the threshold matrix is ​​generated, these upper and lower limits are embedded in the state vector to form a set of dynamic standardized judgment criteria. The generated real-time state vector is compared with the threshold matrix, and each parameter in the state vector is classified level by level through the hierarchical path of GBDT. Specifically, the GBDT model first screens out features that may be abnormal through a hierarchical path, and compares whether the fluctuation amplitude of the parameters meets the classification conditions of "fluctuation" or "fault". During the classification process, the GBDT model will give priority to more sensitive parameters such as current and voltage to ensure that the first step of screening can quickly and accurately identify key anomalies. When a parameter exceeds the set safety threshold range, the system will further trigger the determination of the "fluctuation" or "fault" state. For example, if the current value deviates from the normal range but does not reach the severity, the system will determine the state as "fluctuation"; when the current deviation exceeds the tolerance and persists, it is determined as a "fault" and immediately marked for in-depth processing by the subsequent analysis module.

[0026] Furthermore, the gradient boosting decision tree is constructed by the following steps: Clean and normalize the historical high-dimensional feature matrix set; Based on the parameters in the feature matrix, the key parameters that have a significant impact on anomaly classification are screened out through the SHAP value; Based on the selected key parameters, an initial decision tree model is constructed with the goal of minimizing the classification error, and the initial branch path is set according to the distribution characteristics of the parameters under different abnormal conditions; In each training iteration, the sample weights are adjusted according to the classification error of the previous round, and a new decision tree is constructed through the residual minimization strategy. The new decision tree is combined with the previous decision tree to generate a gradient boosting decision tree model.

[0027] In some embodiments, first, the historical high-dimensional feature matrix set is cleaned and normalized. The feature matrix set contains multi-dimensional parameters, such as current, voltage, temperature and humidity, etc. Each parameter may be affected by noise and produce abnormal values ​​in different time periods. During the cleaning process, the system uses outlier detection methods (such as the 3-times standard deviation method) to identify and remove values ​​that do not conform to the conventional range, thereby improving the stability of the data. After the cleaning is completed, each feature parameter in the matrix is ​​normalized so that the data is mapped between 0 and 1 to ensure that the scales of different parameters in the model are consistent, providing basic support for the accuracy of subsequent models. After data preprocessing, the system screens the parameters in the high-dimensional feature matrix based on the feature importance algorithm SHAP (SHapley AdditiveexPlanations) to determine the key parameters that have a significant impact on abnormal classification. During the SHAP value calculation process, the system evaluates the marginal contribution of each feature to the classification result one by one, and gives higher weights to those features that have a greater impact on the model prediction results. For example, in the abnormal classification of the substation unit, current and insulation resistance may have a higher impact weight, while parameters such as temperature and humidity have a smaller impact. Through the analysis of SHAP values, the system screens out key parameters with high importance, thereby effectively reducing the feature space, reducing the computational complexity of the model, and improving the sensitivity of the model to abnormal conditions. Based on the screened key parameters, the system constructs an initial decision tree model and sets the initial branch path with the goal of minimizing the classification error. When constructing the initial decision tree, the system divides the training data set into multiple layers of nodes and constructs branches based on the distribution characteristics of key parameters under different abnormal conditions. For example, when the current exceeds a certain threshold but the temperature and humidity are still within the normal range, the system classifies it as a "fluctuation" state; if both the current and the insulation resistance exceed the threshold, it is classified as a "fault" state. The division of each branch node is based on indicators such as information gain or Gini index to ensure that the data is segmented in the best way at each layer of nodes, thereby improving the classification accuracy of the initial model. After completing the construction of the initial decision tree, the system enters the iterative training stage, with the residual minimization strategy as the core, and gradually optimizes the model. In each iteration, the system adjusts the weight of the sample based on the classification error of the previous round, and gives higher weights to samples with incorrect classification to increase the model's attention in subsequent decision trees. Based on the adjusted sample weights, the system constructs a new decision tree to supplement the samples that the previous decision tree failed to correctly classify. In this process, the new decision tree is combined with the previous decision tree layer by layer through the residual minimization strategy to ensure that the model can more accurately classify abnormal features in the data set. The overall output of the model is obtained by the integrated weighted average of multiple decision trees, and finally a complete gradient boosting decision tree model is formed.

[0028] Furthermore, the formula of the gradient boosting decision tree is as follows: ; in, is the prediction model of the mth iteration; is the prediction model of the previous round; is the learning rate; is the number of samples in the training data set; is the i-th sample; is the i-th sample The optimal weight corresponding to the residual of ; Weak decision tree generated for this round.

[0029] Specifically, we start with an initial model, gradually reduce the prediction error, and add a new weak decision tree in each iteration to improve the model's fitting effect on the sample. In each iteration, the new decision tree focuses on fitting the error part of the previous model, thereby gradually optimizing the accuracy of the overall model. First, the system starts with a simple initial model, which usually uses the average value of the target variable in the training set as the predicted value. Since this initial model cannot accurately predict in most cases, it will produce a certain error. The goal of GBDT is to reduce this error through iterations so that the model continues to approach the true value. In each iteration, the system calculates the error of the current model for all samples, that is, the difference between the true value and the predicted value of each sample. This difference is called the residual. In each round, the residual reflects the features that the current model fails to accurately capture, so the system builds a new decision tree based on these residuals. The new decision tree is called a "weak decision tree", which is specifically used to fit these residuals to make up for the shortcomings of the current model. In order to control the learning process of the model and prevent overfitting, the system introduces a parameter called "learning rate". The role of the learning rate is to control the influence of each new decision tree on the model. By introducing a learning rate into the model, the system can ensure that the model is properly adjusted in each round of iteration without causing large fluctuations in the prediction results. Each newly generated decision tree is given a weight when it is added to the model, which is determined by the optimization strategy of minimizing the residual. This means that the system will give higher weights to samples with large errors at the moment, ensuring that more attention is paid to these difficult-to-fit samples in the next round of decision trees. Therefore, the GBDT model gradually forms a more accurate prediction model by continuously accumulating the results of each round of weak decision trees. The final GBDT model is the weighted cumulative result of the weak decision trees generated during all iterations. In testing or actual applications, when new data is input, the system will calculate through all decision trees and accumulate the results of each tree to generate the final prediction result.

[0030] Furthermore, the feature extraction of the state vector using a variational autoencoder and a self-attention mechanism according to the abnormal state includes the following steps: The state vector is input into the encoding layer of the variational autoencoder, and the state vector is encoded with high-dimensional features through several layers of neural networks to generate latent variable representation; Apply the self-attention mechanism to the latent variable representation of the variational autoencoder output and calculate the attention weight of each dimension feature; The feature representation processed by the self-attention mechanism is input into the decoding layer of the variational autoencoder to generate a reconstructed state vector; The difference between the reconstructed state vector and the original state vector is calculated to determine the abnormal feature vector, and the corresponding potential fault is matched according to the abnormal feature vector.

[0031] Specifically, the intelligent control system of the substation uses a variational autoencoder (VAE) and a self-attention mechanism to extract features from the state vector of the substation unit, thereby identifying abnormal states and extracting possible fault features. This process uses a multi-layer neural network to implement high-dimensional feature encoding, feature weight assignment, and reconstruction difference calculation to determine the abnormal state. First, the state vector is input into the encoding layer of the VAE. The encoder of the VAE consists of several layers of neural networks, which are used to compress the input high-dimensional state vector and extract its potential features. During the encoding process, the system extracts high-dimensional features from the state vector layer by layer according to the weights and activation functions of each layer, and compresses it into a latent variable representation. The latent variable representation contains the key features of the state vector while retaining the statistical information of the original state vector, providing a concise and effective representation for subsequent feature analysis. Taking multiple monitoring parameters such as current and voltage as an example, the encoding layer captures their joint characteristics in a specific state by high-dimensional compression of these parameters, so that the latent variable can more effectively express the correlation between these features. Next, the self-attention mechanism is applied to the latent variable representation of the VAE encoding output to calculate the importance weight of each dimensional feature. The role of the self-attention mechanism is to highlight the features that have a key impact on fault detection while reducing the impact of other unimportant features. The self-attention mechanism calculates the attention weights between different feature dimensions by constructing a query, key, and value matrix. Specifically, the system first matches the latent variable representation with itself for similarity, and calculates the attention weight matrix based on the similarity between each dimension. For key parameters such as current and voltage in the substation unit, the self-attention mechanism automatically assigns higher weights to these features to ensure that the impact of these parameters is highlighted in subsequent analysis. For example, in an abnormal state where the current changes significantly, the self-attention mechanism assigns higher weights to the current features, allowing the system to focus on these key parameters. Subsequently, the feature representation processed by the self-attention mechanism is input into the decoding layer of the VAE to generate a reconstructed state vector. The decoding layer reconstructs the original state by gradually restoring the feature vector to a space consistent with the input dimension. The decoding layer has a similar structure to the encoding layer. By reversely restoring layer by layer, the encoded compressed information is restored to an approximate representation of the original state. Through this process, the system is able to generate a reconstructed state vector whose value is close to the original state vector, but the feature structure is adjusted by the self-attention mechanism to highlight the key features. Finally, the system calculates the difference between the reconstructed state vector and the original state vector to determine the abnormal feature vector. The difference calculation usually uses the mean square error (MSE) or similarity index to indicate the degree of deviation between the reconstructed state and the original state. The size of the difference value directly reflects the system's judgment on the normality of the current state. The larger the difference, the higher the possibility of abnormality.For example, under normal circumstances, the difference between the reconstructed vector and the original vector should be small; if there is a significant difference, the system will mark the difference as an abnormal feature vector. The abnormal feature vector is compared with the preset fault template or historical data to determine the potential fault type that matches it. In this way, the system can not only effectively identify the current abnormal state, but also preliminarily locate the possible cause of the fault, thus providing an important basis for subsequent fault handling.

[0032] Furthermore, the method of predicting future power demand data based on the long short-term memory network through historical substation load data and real-time substation load data includes the following steps: De-noising, smoothing and normalizing the historical load data and real-time load data, converting them into a time series data matrix, and constructing time windows in segments to generate a standardized load data input matrix; The standardized load data input matrix is ​​divided into blocks according to the time dimension, and the load amount and timestamp are mapped into a multi-dimensional feature vector sequence; Add time labels and load change rates to the feature sequences in each time window to form a time dependency matrix as the input of the long short-term memory network model; The time dependency matrix is ​​input into the long short-term memory network model, and forward and backward propagation training is performed on each input time window, and the model parameters are optimized by gradient descent; Input the real-time load data sequence into the trained long short-term memory network model, and use the memory unit to update the state of the sequence features; The future electricity demand data is output through the output layer.

[0033] In some embodiments, first, the historical load data and real-time load data are denoised, smoothed and normalized. The denoising operation removes high-frequency noise in the data through a filtering algorithm to eliminate the impact of abnormal fluctuations on load forecasting; the smoothing process uses a moving average or weighted average method to make the data curve more stable, which is convenient for the model to capture long-term trends. Subsequently, the normalization operation converts the data to a uniform scale range (usually between 0 and 1) to ensure that data of different magnitudes have consistent numerical weights in the model. The processed data is organized into a time series data matrix, and the time window is segmented at fixed time intervals to generate a standardized load data input matrix to provide a standardized data structure for subsequent model training. Based on the generated standardized load data input matrix, the data is divided into blocks according to the time dimension. This step maps the load amount and timestamp into a multi-dimensional feature vector sequence, so that LSTM can capture the time dependency of load changes. Through block processing, the load data in the time series is converted into a structured input that can be understood by LSTM. The block data not only retains the continuity of the load data, but also provides it to the model in a segmented form, thereby improving the flexibility and accuracy of the prediction. On the basis of the feature vector sequence generated by blocks, additional information such as time labels and load change rate are added to the feature sequence of each time window to generate a time dependency matrix. The time label indicates the relative position of each time window in the sequence, so that the model can more effectively understand the time series relationship between short-term and long-term loads; the load change rate reflects the speed and trend of change between the current load and the previous load, providing auxiliary information for the model to identify load peaks and valleys. The time dependency matrix is ​​used as the input matrix of the LSTM model to ensure that the model can receive multi-dimensional time series features during training. The time dependency matrix is ​​input into the LSTM model, and the model begins to perform forward and backward propagation training on the input of each time window. In this process, the "memory unit" of the LSTM will save state information based on the data of the previous time window to capture the long-term trend of load changes; at the same time, the "forget gate" mechanism is used to discard useless information at the appropriate time to ensure that the model can dynamically adjust the data transmission and memory state. Through forward propagation, the model updates the state according to the current input, and in backward propagation, the model optimizes parameters based on the gradient descent method to gradually reduce the prediction error, thereby improving the accuracy and generalization ability of the model. After the model training is completed, the real-time load data sequence is input into the trained LSTM model to update the state of the memory unit. The LSTM model can rely on its "long short-term memory" mechanism in each time step to adjust the state of the load data in real time, and gradually accumulate the time series characteristics of the data, so that the prediction can more accurately reflect the changing trend of the real-time load. Finally, through the output layer of LSTM, the model outputs future power demand data. The output results include the predicted load at multiple time points in the future. Through multiple rounds of iterations, the model combines short-term memory with long-term memory to generate accurate prediction values ​​for future loads.This prediction data can be used as input for the intelligent control module to support the dynamic allocation of loads and resource scheduling of substations, thereby improving the response speed and efficiency of the system in actual operation. Through this LSTM model prediction process, the system can adapt to fluctuations in power demand in real time and accurately, and achieve optimized management of power resources.

[0034] Furthermore, the steps of constructing the reinforcement learning model include: The predicted future power demand data is used as the environmental input state of the reinforcement learning model to initialize the operating state and demand load of each substation unit in the substation to form the initial state matrix of the system; According to the operation status of the substation unit, a set of control actions is constructed, including increasing, decreasing or maintaining the number of substation units in operation; Construct the reward function so that the number of runs is negatively correlated with the reward value in the reinforcement learning model; By making strategic decisions based on the current state matrix and reward function in each time step, the action with the highest reward value is selected to adjust the number of substation operations; The load response, reward value, and state change data after each action are recorded and fed back to the model to update the state matrix and optimize the reinforcement learning model.

[0035] In some embodiments, first, the predicted future power demand data is used as the environmental input state of the reinforcement learning model, and the current operating state and required load of each substation in the substation are initialized. The system organizes the initial parameters of each substation (such as current load, power, current, etc.) and the predicted load demand into the initial state matrix of the system. This state matrix represents the overall operating state of the current system, including the working state and load demand of each substation, which provides a complete environmental description for the reinforcement learning model, enabling the model to understand the operation of each substation under a given load demand. For the operating state of the substation, the system constructs a set of control actions to cope with the fluctuation of load demand. The control action set includes increasing, decreasing or maintaining the operating number of substations, which is specifically defined as: increasing the operating number means starting additional substations to cope with future load increases; reducing the operating number means shutting down some substations to save energy; maintaining the operating number is used to maintain the current working state when the load demand is stable. This action set provides operation options for the reinforcement learning model, enabling the system to be flexibly adjusted under different load conditions. To ensure that the model can balance efficient operation and resource conservation when operating the substation, the system constructs a reward function and makes the number of operations negatively correlated with the reward value. The reward function considers the impact of the number of operations on resource consumption. Increasing the number of operations will lead to a decrease in the reward value, while reducing the number of operations will increase the reward value. In this way, the model will tend to select fewer substations while meeting the load demand to optimize resource utilization. For example, when the load demand is peaking, the model will balance the load demand and resource cost through the reward function and minimize the number of operations while meeting the demand. In actual operation, the system makes policy decisions based on the current state matrix and the reward function at each time step, selects the action with the highest reward value, and adjusts the number of substation operations. At each time step, the reinforcement learning model calculates the reward value corresponding to each action based on the current system state and future load forecasts, and selects the action that can maximize the reward to ensure that the system can achieve optimal control at a future time point. For example, if the forecast shows that the load is about to rise, the model will moderately increase the number of substation operations; when the load decreases, the model will choose to reduce the number of operations to save resources. After each action is executed, the system records the load response, reward value, and state change data after the action is executed, and feeds this data back to the reinforcement learning model to update the state matrix and optimize the model's strategy. In this feedback process, the model updates the strategy parameters to make future decisions more accurate and efficient, thereby gradually improving the model's intelligence and adaptability. This process continues, allowing the reinforcement learning model to continuously adapt to the actual load requirements of the substation while also accumulating and optimizing its control strategy to achieve optimal resource allocation and efficient operation.

[0036] The above implementation modes are merely descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering and technical personnel in the field shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A standardized intelligent control system for a substation, characterized in that: include: A distributed monitoring module, a fault detection module, a load monitoring module and an intelligent control module, wherein the distributed monitoring module and the fault detection module are connected in sequence, and the load monitoring module and the intelligent control module are connected; The distributed monitoring module is used to obtain active parameters and passive parameters of the substation units in the substation; The fault detection module is used to receive active parameters and passive parameters of a single substation and standardize them to generate a substation state vector, perform preliminary abnormal classification based on a gradient boosting decision tree, and determine the abnormal state; according to the abnormal state, use a variational autoencoder and a self-attention mechanism to extract features from the state vector and detect potential faults; The load monitoring module is used to obtain and store substation load data; The intelligent control module is used to predict future power demand data through historical substation load data and real-time substation load data based on the long short-term memory network, and control the operating quantity of substation units without abnormalities and potential faults through a reinforcement learning model based on the future power demand data, and the operating quantity is negatively correlated with the reward value in the reinforcement learning model; The receiving active parameters of a single substation unit and standardizing and generating a substation unit state vector, and performing preliminary abnormality classification based on a gradient boosting decision tree include: Active parameters and passive parameters are fused to generate a high-dimensional feature matrix, and a real-time dynamic state vector is generated based on the matrix; According to the historical high-dimensional feature matrix set, a threshold matrix is ​​generated, and the adaptive threshold of each parameter is embedded in the state vector; The state vector is compared with the threshold matrix and classified based on the hierarchical path of the gradient boosting decision tree.

2. A standardized intelligent control system for a substation according to claim 1, characterized in that: The substation unit includes a main transformer, a high-voltage switch cabinet, a low-voltage switch cabinet, a circuit breaker, a busbar system, a cable connector and a lightning arrester.

3. A standardized intelligent control system for a substation according to claim 1, characterized in that: The distributed monitoring module includes a current sensor, a voltage sensor, a temperature and humidity sensor, an insulation detector and an embedded operation monitoring system.

4. A standardized intelligent control system for a substation according to claim 1, characterized in that: The receiving active parameters of a single substation unit and standardizing and generating a substation unit state vector, and performing preliminary abnormal classification based on a gradient boosting decision tree specifically include the following steps: De-noise the active and passive parameters and fuse them to generate a high-dimensional feature matrix; Based on the high-dimensional feature matrix, an adaptive normalization model is used to generate real-time dynamic state vectors; Generate a threshold matrix based on the historical high-dimensional feature matrix set through the gradient boosting decision tree, and embed the adaptive threshold of each parameter into the state vector; The state vector is compared with the threshold matrix, and the state vector is classified based on the hierarchical path of the gradient boosting decision tree to determine the abnormal state, which includes no abnormality, fluctuation and failure.

5. A standardized intelligent control system for a substation according to claim 4, characterized in that: The gradient boosting decision tree is constructed by the following steps: Clean and normalize the historical high-dimensional feature matrix set; Based on the parameters in the feature matrix, the key parameters that have a significant impact on anomaly classification are screened out through the SHAP value; Based on the selected key parameters, an initial decision tree model is constructed with the goal of minimizing the classification error, and the initial branch path is set according to the distribution characteristics of the parameters under different abnormal conditions; In each training iteration, the sample weights are adjusted according to the classification error of the previous round, and a new decision tree is constructed through the residual minimization strategy. The new decision tree is combined with the previous decision tree to generate a gradient boosting decision tree model.

6. A standardized intelligent control system for a substation according to claim 5, characterized in that: The formula for the gradient boosting decision tree is as follows: ； in, is the prediction model of the mth iteration; is the prediction model of the previous round; is the learning rate; is the number of samples in the training data set; is the i-th sample; is the i-th sample The optimal weight corresponding to the residual of ; Weak decision tree generated for this round.

7. A standardized intelligent control system for a substation according to claim 5, characterized in that: The method of extracting features from the state vector using a variational autoencoder and a self-attention mechanism according to the abnormal state includes the following steps: The state vector is input into the encoding layer of the variational autoencoder, and the state vector is encoded with high-dimensional features through several layers of neural networks to generate latent variable representation; Apply the self-attention mechanism to the latent variable representation of the variational autoencoder output and calculate the attention weight of each dimension feature; The feature representation processed by the self-attention mechanism is input into the decoding layer of the variational autoencoder to generate a reconstructed state vector; The difference between the reconstructed state vector and the original state vector is calculated to determine the abnormal feature vector, and the corresponding potential fault is matched according to the abnormal feature vector.

8. A standardized intelligent control system for a substation according to claim 1, characterized in that: The method of predicting future power demand data based on the long short-term memory network by using historical substation load data and real-time substation load data includes the following steps: De-noising, smoothing and normalizing the historical load data and real-time load data, converting them into a time series data matrix, and constructing time windows in segments to generate a standardized load data input matrix; The standardized load data input matrix is ​​divided into blocks according to the time dimension, and the load amount and timestamp are mapped into a multi-dimensional feature vector sequence; Add time labels and load change rates to the feature sequences in each time window to form a time dependency matrix as the input of the long short-term memory network model; The time dependency matrix is ​​input into the long short-term memory network model, and forward and backward propagation training is performed on each input time window, and the model parameters are optimized by gradient descent; Input the real-time load data sequence into the trained long short-term memory network model, and use the memory unit to update the state of the sequence features; The future electricity demand data is output through the output layer.

9. A standardized intelligent control system for a substation according to claim 8, characterized in that: The steps of constructing the reinforcement learning model include: The predicted future power demand data is used as the environmental input state of the reinforcement learning model to initialize the operating state and demand load of each substation unit in the substation to form the initial state matrix of the system; According to the operation status of the substation unit, a set of control actions is constructed, including increasing, decreasing or maintaining the number of substation units in operation; Construct the reward function so that the number of runs is negatively correlated with the reward value in the reinforcement learning model; By making strategic decisions based on the current state matrix and reward function in each time step, the action with the highest reward value is selected to adjust the number of substation operations; The load response, reward value, and state change data after each action are recorded and fed back to the model to update the state matrix and optimize the reinforcement learning model.

10. A standardized intelligent control system for a substation according to claim 1, characterized in that: The active parameters include current, voltage, power factor, temperature and humidity, operating status and insulation status, and the passive parameters include ambient temperature and humidity inside and outside the substation.

Images