Power utilization safety monitoring method for power distribution station

Abstract

The invention provides a power distribution station power utilization safety monitoring method, which relates to the field of power distribution station monitoring and comprises the following steps: acquiring electrical parameters, environmental parameters and video data in a power distribution station, and uploading the acquired data to a control end; the control end obtains training data by using a data protocol conversion program of the Node-RED; constructing a monitoring model based on artificial intelligence, and training the monitoring model by using the training data, the monitoring model comprises a shared feature extractor, a security risk prediction branch, an anomaly prediction branch and an energy efficiency analysis branch which are connected with the shared feature extractor, and a fusion module which fuses and decides results obtained by the security risk prediction branch, the anomaly prediction branch and the energy efficiency analysis branch; and triggering a corresponding alarm level according to the fusion module output. The intelligent monitoring system can be compatible with multi-protocol equipment, realizes multi-source data fusion, and performs comprehensive analysis based on artificial intelligence so as to improve the safety and management efficiency of the power distribution station.

Description

Technical Field

[0001] This invention relates to the field of power distribution station monitoring, and in particular to a method for monitoring the safety of electricity use in power distribution stations. Background Technology

[0002] Substations are critical nodes connecting the main grid and users in a power system, and their operational status directly affects power supply reliability and electricity safety. However, current substation monitoring systems generally suffer from the following shortcomings: Limited monitoring dimensions: most systems only collect basic electrical parameters such as current and voltage, lacking comprehensive awareness of environmental factors (such as temperature, humidity, and water immersion), equipment status (such as switch status), and video information; incompatible communication protocols: different manufacturers and types of equipment often use different communication protocols (such as Modbus, MQTT, CoAP, and HTTP), making it difficult for traditional systems to access them uniformly, resulting in data silos; low level of intelligence: existing systems mostly rely on fixed thresholds for simple alarms, lacking the ability to fuse and analyze multi-dimensional data and make intelligent judgments, making it impossible to predict potential faults or conduct energy efficiency assessments in advance; and high reliance on manual intervention: anomaly detection and maintenance decisions mainly depend on human experience, resulting in low efficiency and slow response. Summary of the Invention

[0003] The main objective of this invention is to propose a method for monitoring the power safety of power distribution stations, which is an intelligent monitoring system that is compatible with multi-protocol devices, realizes multi-source data fusion, and performs comprehensive analysis based on artificial intelligence, so as to improve the safety and management efficiency of power distribution stations.

[0004] This invention is achieved through the following technical solution:

[0005] The method for monitoring the safety of electricity use in power distribution stations includes the following steps:

[0006] Step S1: Collect electrical parameters, environmental parameters, and video data within the substation. Using the DTU data transmission unit, upload the collected data to the control terminal via the communication protocol corresponding to the data type.

[0007] Step S2: The control end uses the Node-RED data protocol conversion program to process the data uploaded using different communication protocols to obtain training data in a unified format.

[0008] Step S3: Construct an artificial intelligence-based monitoring model and train the monitoring model using training data. The monitoring model includes a shared feature extractor, a safety risk prediction branch, an anomaly prediction branch, and an energy efficiency analysis branch connected to the shared feature extractor, and a fusion module that integrates and makes decisions based on the results obtained from the safety risk prediction branch, anomaly prediction branch, and energy efficiency analysis branch.

[0009] Step S4: Trigger the corresponding alarm level based on the output of the fusion module.

[0010] Furthermore, in step S1, the substation is equipped with current sensors, voltage sensors and / or power sensors for collecting electrical parameters, temperature and humidity sensors and / or water immersion sensors for collecting environmental parameters, and video monitoring equipment for collecting video data. Each sensor and video monitoring equipment is connected to the DTU data transmission unit, and the communication protocol includes Modbus, MQTT, CoAP or HTTP.

[0011] Furthermore, in step S3, the shared feature extraction module adopts a multi-layer neural network structure, including a feature embedding layer, a temporal feature extraction layer, a high-level feature abstraction layer, and an output layer. Both the feature embedding layer and the temporal feature extraction layer adopt neural network structures. The training data X is processed by the feature embedding layer to obtain h1=σ(W1·X+b1), and h1 is processed by the bidirectional LSTM of the temporal feature extraction layer to capture temporal features. After h2 is processed by the high-level feature abstraction layer, h3 = σ(W3·h2 + b3) is obtained. After h3 is normalized by the output layer, the shared feature F = LayerNorm(h3) is obtained. , W1 and W3 are the weight matrices of the feature embedding layer and the high-level feature abstraction layer, respectively, b1 and b3 are the bias vectors of the feature embedding layer and the high-level feature abstraction layer, respectively, and σ represents the ReLU activation function.

[0012] Furthermore, in step S3, the security risk prediction branch includes a risk input layer, a first temporal modeling layer, an attention mechanism layer, a security risk prediction layer, and a first output layer. The risk input layer is based on F. risk =W risk ·F+b risk Extract risk-related features F from shared features F. risk The first temporal modeling layer uses two LSTM layers for risk temporal prediction, resulting in r2 = LSTM2(r1). The attention mechanism layer introduces temporal attention weights α = softmax(W). α ·r2+b α ), to obtain r att =Σ(α⊙r2), the security risk prediction layer is based on y risk =softmax(W p ·r att +b p ) to predict the risk level, and based on p risk =sigmoid(U p ·r att +c p ) Perform risk probability calculation, and the first output layer outputs RiskOutput=(y risk ,prisk ,t), where r1=LSTM1(F risk ), LSTM1 represents the first LSTM layer, LSTM2 represents the second LSTM layer, ⊙ represents element-wise multiplication, W α b α W represents the weight matrix and bias vector of the attention mechanism layer. p b p U represents the weight matrix and bias vector for risk level prediction. p c p This represents the weight matrix and bias vector used in the risk probability calculation, where t is the timestamp for the safety risk prediction, and y represents the risk level. risk ∈ , Let K be the set of real numbers, and K be the set number of risk levels.

[0013] Furthermore, in step S3, the anomaly prediction branch includes an anomaly input layer, a convolutional feature extraction layer, a second temporal modeling layer, an anomaly scoring calculation layer, an anomaly type classification layer, and a second output layer. The anomaly input layer is based on F... anomaly =W anom ·F+b anom Extract anomaly-related features F from shared features F. anomaly The convolutional feature extraction layer uses one-dimensional convolution to capture local anomaly patterns, resulting in c2=MaxPooling(c1) and c1=Conv1D(F). anomaly kernel size =3), the second time series modeling layer uses two bidirectional LSTM layers to obtain b2=BiLSTM2(b1), b1=BiLSTM1(c2), and the anomaly scoring calculation layer is based on formula p anomaly =sigmoid(W s ·flatten(b2)+b s ) Calculate the anomaly score; the anomaly type classification layer is based on y type =softmax(W t ·b2[:,-1,:]+b t Predict the probability distribution of anomaly types; the second output layer outputs AnomalyOutput=(p anomaly ,y type ,t), where BiLSTM1 represents the first bidirectional LSTM layer, BiLSTM2 represents the second bidirectional LSTM layer, and W s b s Let b1 represent the weight matrix and bias vector of the anomaly scoring calculation layer, b2[:,-1,:] represent the hidden state taken from the last time step of the bidirectional LSTM layer output, and p... anomaly ∈[0,1].

[0014] Furthermore, the energy efficiency analysis branch includes an energy efficiency input layer, a self-attention mechanism layer, an energy efficiency index layer, an energy efficiency scoring layer, and a third output layer. The energy efficiency input layer is based on F... eff =W eff ·F+b eff Extracting energy efficiency-related features F eff Energy efficiency related characteristics F eff Inputting into the self-attention mechanism layer, we obtain F att =A·V, Power Factor Efficiency (PF) is calculated using the energy efficiency index layer. efficiency =MLP(F att [:,t pf ,:]) and load balancing balance =MLP(mean(F att The energy efficiency rating layer obtains E score =w1·PF efficiency +w2·Load balance +b, the third output layer outputs EfficiencyOutput=(PF efficiency Load balance E score ,t), where A=softmax(Q·Kᵀ / ), Q=W Q ·F eff K=W K ·F eff V=W V ·F eff W eff b eff This represents the weight matrix and bias vector of the energy-efficient input layer, where d is the dimension of K, MLP represents a multilayer perceptron, and t pf The key time step index, F, represents the power factor. att [:,t pf ,:] indicates the critical sampling time t for the power factor. pf eigenvectors, mean(F) att This indicates that Fatt is averaged over time, where w1 and w2 are weighting coefficients, b is the bias term for the energy efficiency score, and W... Q W K W V These represent the transformation weight matrices for Query, Key, and Value in the self-attention mechanism, respectively.

[0015] Furthermore, the fusion module utilizes DS evidence theory to fuse the outputs of the safety risk prediction branch, the anomaly prediction branch, and the energy efficiency analysis branch.

[0016] Furthermore, the loss function L is adopted.tota =λ1L_ risk +λ2L_ anomaly +λ3L_ efficiency The monitoring model is trained using an alternating training strategy, where L risk =CrossEntropy(y risk ,ŷ risk )+MSE(p risk ,p̂ risk ), L anomaly =BCE(p anomaly ,p̂ anomaly )+CrossEntropy(y type ,ŷ type ), L efficiency =MSE(PF efficiency ,PF̂ efficiency )+MSE(Load balance L̂oad balance )+MSE(E score ,Ê score ), ŷ risk The true label indicating the risk level, p̂ risk The true label representing the probability of risk, p̂ anomaly For the true label of outlier scores, ŷ type For the true label of the exception type, PF̂ efficiency For the true value of power factor efficiency, L̂oad balance For the true value of load balancing, Ê score For the true value of the energy efficiency score, CrossEntropy(.) represents the cross-entropy loss function, MSE(.) represents the mean squared error loss function, BCE(.) represents the binary cross-entropy loss function, and λ1, λ2, and λ3 are the weighting coefficients of the three branch loss functions.

[0017] As can be seen from the above description of the present invention, compared with the prior art, the present invention has the following beneficial effects:

[0018] This invention first collects electrical parameters, environmental parameters, and video data within the substation. Using a DTU (Distributed Terminal Unit) data transmission unit, the collected data is uploaded to the control terminal via communication protocols corresponding to the data types. The control terminal then uses a Node-RED data protocol conversion program to process the data uploaded using different communication protocols, obtaining training data in a unified format. Next, an AI-based monitoring model is constructed and trained using the training data. The monitoring model includes a shared feature extractor, a safety risk prediction branch, an anomaly prediction branch, and an energy efficiency analysis branch connected to the shared feature extractor, and a fusion module that integrates and makes decisions based on the results from the safety risk prediction branch, anomaly prediction branch, and energy efficiency analysis branch. Finally, the corresponding alarm level is triggered based on the output of the fusion module. This enables multi-dimensional data collection, making monitoring more comprehensive, flexibly supporting multiple communication protocols, improving system compatibility and scalability, and using the monitoring model for prediction to achieve predictive maintenance, upgrading from traditional "passive response" to "proactive prevention," significantly improving the operational safety and maintenance efficiency of the substation. Attached Figure Description

[0019] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] Figure 1 This is a flowchart of the present invention.

[0021] Figure 2 This is a flowchart illustrating the protocol conversion process based on Node-RED in this invention.

[0022] Figure 3 This is a schematic diagram of the monitoring model of the present invention. Detailed Implementation

[0023] The present invention will be further described below through specific embodiments.

[0024] like Figure 1 As shown, the method for monitoring the safety of electricity use in a power distribution station includes the following steps:

[0025] Step S1: Collect electrical parameters, environmental parameters, and video data within the substation. Using the DTU data transmission unit, upload the collected data to the control terminal via the communication protocol corresponding to the data type.

[0026] The substation is equipped with current sensors, voltage sensors, and / or power sensors to collect electrical parameters; temperature and humidity sensors and / or water immersion sensors to collect environmental parameters; and video surveillance equipment to collect video data. All sensors and video surveillance equipment are connected to the DTU data transmission unit via RS485, LoRa, Wi-Fi, etc. The control terminal communication protocols include Modbus, MQTT, CoAP, or HTTP. The characteristics of each communication protocol are adapted as follows:

[0027] Modbus RTU / TCP: For power meters (such as Schneider PM800), a register address mapping table needs to be defined (e.g., hold registers 40001-40005 for voltage / current / power factor).

[0028] MQTT protocol: Configure QoS level to 1 (at least one transmission), and design the topic subscription structure as / substation / {device ID} / sensor / {data type};

[0029] CoAP protocol: Adapts to DLMS / COSEM standards and processes OBIS encoded data;

[0030] HTTP / HTTPS: Supports JSON format data exchange and sets API authentication (JWT token verification).

[0031] Step S2: The control end uses the Node-RED data protocol conversion program to process the data uploaded using different communication protocols to obtain training data in a unified format.

[0032] like Figure 2 As shown, the process of parsing, mapping and standardizing multi-protocol data (such as Modbus, MQTT, HTTP, CoAP, etc.) through Node-RED enables unified access and data formatting for devices with different communication protocols, providing standardized input for subsequent monitoring models.

[0033] More specifically, the node-red-contrib-modbus node is deployed in Node-RED to process Modbus data, which is implemented through function nodes; a custom Node-RED node protocol-mapper is developed to automatically load parsing rules based on device metadata; and a unified JSON format data is output as training data.

[0034] Step S3: Construct an artificial intelligence-based monitoring model and train the monitoring model using training data. The monitoring model includes a shared feature extractor, a safety risk prediction branch, an anomaly prediction branch, and an energy efficiency analysis branch connected to the shared feature extractor, and a fusion module that integrates and makes decisions based on the results obtained from the safety risk prediction branch, anomaly prediction branch, and energy efficiency analysis branch.

[0035] The architecture diagram of the monitoring model is as follows: Figure 3 As shown. The shared feature extraction module adopts a multi-layer neural network structure, including a feature embedding layer, a temporal feature extraction layer, a high-level feature abstraction layer, and an output layer. Training data... After the feature embedding layer, we obtain h1 = σ(W1·X + b1). h1 then passes through a bidirectional LSTM layer in the temporal feature extraction layer to capture temporal features. h2 is obtained after being processed by a high-level feature abstraction layer. After h3 is normalized by the output layer, the shared features are obtained. ,in, Let T represent the set of real numbers, T represent the time step (e.g., one data point per hour, T=24 represents one day's data), and D=D1+D2+D3 represent the feature dimensions. D1 dimension refers to electrical parameters (current I, voltage V, power P, power factor PF, etc.), D2 dimension refers to environmental parameters (temperature Temp, humidity Hum, water immersion status WS, etc.), and D3 dimension refers to video features (equipment status features VS, personnel behavior features PA, etc.). This is the weight matrix for the feature embedding layer, which consists of D rows and H1 columns. Here, σ represents the bias vector of the feature embedding layer, σ denotes the ReLU activation function, and H1 is the number of hidden units in the feature embedding layer. , H2 represents the number of hidden units in the temporal feature extraction layer. W3 is the number of hidden units in the high-level feature abstraction layer, W3 is the weight matrix of the high-level feature abstraction layer, b3 is the bias vector of the high-level feature abstraction layer, σ represents the ReLU activation function, and LayerNorm is the layer normalization operation to ensure that the input scales of the safety risk prediction branch, the anomaly prediction branch and the energy efficiency analysis branch are consistent.

[0036] The security risk prediction branch includes a risk input layer, a first temporal modeling layer, an attention mechanism layer, a security risk prediction layer, and a first output layer. The risk input layer is based on... Extract risk-related features F from shared features F. risk The first time series modeling layer uses two LSTM layers for risk time series prediction, resulting in... The attention mechanism layer introduces a temporal attention weight α = softmax(W). α ·r2+b α),get Security risk prediction layer Perform risk level prediction and based on p risk =sigmoid(U p ·r att +c p ) Perform risk probability calculation, and the first output layer outputs RiskOutput=(y risk ,p risk ,t), where, , H represents the weight matrix and bias vector of the risk input layer, respectively. r The number of hidden units in the risk input layer. The number of hidden units for the first temporal modeling layer and the attention mechanism layer, r1=LSTM1(F risk ), LSTM1 represents the first LSTM layer, LSTM2 represents the second LSTM layer, ⊙ represents element-wise multiplication to achieve context-dependent attention weighting, W α b α W represents the weight matrix and bias vector of the attention mechanism layer. p b p U represents the weight matrix and bias vector for risk level prediction. p c p This represents the weight matrix and bias vector used in the risk probability calculation, where t is the timestamp for the safety risk prediction, K is the set number of risk levels (e.g., low / medium / high), and p... risk ∈[0,1].

[0037] The anomaly prediction branch includes an anomaly input layer, a convolutional feature extraction layer, a second temporal modeling layer, an anomaly scoring calculation layer, an anomaly type classification layer, and a second output layer. The anomaly input layer is based on... Extract anomaly-related features F from shared features F. anomaly The convolutional feature extraction layer uses one-dimensional convolution to capture local anomaly patterns, resulting in... , The second time-series modeling layer uses two bidirectional LSTM layers to obtain... b1=BiLSTM1(c2), the anomaly scoring calculation layer is based on formula p anomaly =sigmoid(W s ·flatten(b2)+b s Anomaly scores are calculated for ∈ [0.1], and the anomaly type classification layer is based on... Predict the probability distribution of anomaly types, and output AnomalyOutput=(p) from the second output layer. anomaly ,y type,t), where BiLSTM1 represents the first bidirectional LSTM layer, BiLSTM2 represents the second bidirectional LSTM layer, M is the number of anomaly types, and W s b s Let b1 represent the weight matrix and bias vector of the anomaly scoring calculation layer. b2[:,-1,:] represents the hidden state at the last time step of the bidirectional LSTM layer output. This state integrates the forward and backward information of the entire time series and has a shape of [batch_size, hidden_dim]. p anomaly ∈[0,1], where t is the anomaly prediction timestamp.

[0038] The energy efficiency analysis branch includes an energy efficiency input layer, a self-attention mechanism layer, an energy efficiency index layer, an energy efficiency scoring layer, and a third output layer. The energy efficiency input layer is based on... Extracting energy efficiency-related features F eff Energy efficiency related characteristics F eff Input to the self-attention mechanism layer, and get Q=W Q ·F eff K=W K ·F eff V=W V ·F eff A = softmax(Q·Kᵀ / The energy efficiency index layer calculates the power factor efficiency (PF). efficiency =MLP(F att [:,t pf ,:]) and load balancing balance =MLP(mean(F att The energy efficiency rating layer obtains E score =w1·PF efficiency +w2·Load balance +b, the third output layer outputs EfficiencyOutput=(PF efficiency Load balance E score ,t), where W eff b eff The weight matrix and bias vector of the energy-efficient input layer are represented by MLP, which stands for Multilayer Perceptron for nonlinear feature transformation. pf The key time step index, F, represents the power factor. att [:,t pf ,:] indicates the critical sampling time t for the power factor. pf eigenvectors, mean(F) attThe expression represents the average of Fatt over time to obtain a global representation of the sequence. w1 and w2 are weighting coefficients, representing the weights of power factor efficiency and load balancing in the overall score, respectively. b is the bias term for the energy efficiency score. Q W K W V represents the transformation weight matrices of Query, Key, and Value in the self-attention mechanism, respectively; d is the dimension of K, used to scale the attention score to avoid gradient vanishing; and t is the energy efficiency analysis timestamp.

[0039] The fusion module utilizes the DS evidence theory to fuse the outputs of the safety risk prediction branch, the anomaly prediction branch, and the energy efficiency analysis branch.

[0040] The monitoring model is trained using training data to determine the optimal weight matrix and bias vector for each layer. A loss function L is employed. tota =λ1L_ risk +λ2L_ anomaly +λ3L_ efficiency The monitoring model is trained using an alternating training strategy, where L risk =CrossEntropy(y risk ,ŷ risk )+MSE(p risk ,p̂ risk ), L anomaly =BCE(p anomaly ,p̂ anomaly )+CrossEntropy(y type ,ŷ type ), L efficiency =MSE(PF efficiency ,PF̂ efficiency )+MSE(Load balance L̂oad balance )+MSE(E score ,Ê score ), ŷ risk The true label indicating the risk level, p̂ risk The true label representing the probability of risk, p̂ anomaly For the true label of outlier scores, ŷ type For the true label of the exception type, PF̂ efficiency For the true value of power factor efficiency, L̂oad balance For the true value of load balancing, Ê scoreThe true value of the energy efficiency score is represented by CrossEntropy(.), which is used for classification tasks; MSE(.), which is used for mean squared error loss functions; and BCE(.), which is used for binary classification or probabilistic prediction tasks. λ1, λ2, and λ3 are the weighting coefficients of the three branch loss functions, which control the relative importance of each task in the total loss.

[0041] The alternating training strategy is as follows:

[0042] 1. Fixed shared parameters, separate training for each task branch;

[0043] 2. Jointly fine-tune all network parameters;

[0044] 3. Dynamically adjust the loss weight λ;

[0045] The AdamW optimizer is used in model optimization, with a learning rate scheduling policy of lr=lr. base *min(step⁻ 05 Step-warmup⁻¹ 5 ).

[0046] Step S4: Trigger the corresponding alarm level based on the output of the fusion module;

[0047] Alarm levels include low, medium, and high. The warning threshold for the substation is dynamically adjusted based on historical data: Threshold = μ + k·σ, where μ is the moving average, σ is the sliding standard deviation, and k is an adjustable parameter.

[0048] To verify the practical effectiveness of this invention, a three-month comparative test was conducted at a 10kV substation in an industrial park, covering the following scenarios:

[0049] Equipment types: Transformers (2 units), high-voltage switchgear (4 sides), low-voltage distribution boxes (8 units), environmental sensors (temperature and humidity / water immersion, 12 units in total);

[0050] Data types: Electrical parameters (current / voltage / temperature, sampling frequency 1Hz), environmental parameters (temperature and humidity / water immersion, sampling frequency 0.5Hz), video data (equipment status / personal behavior, frame rate 30fps);

[0051] Abnormal event database: contains historical fault records (overload, short circuit, equipment overheating, water immersion) and interference events (rainstorms causing a brief increase in temperature and humidity, maintenance personnel touching equipment during temporary maintenance).

[0052] The experimental methods and evaluation indicators are shown in Table 1:

[0053] Table 1

[0054]

[0055] Evaluation metric: False alarm rate = (Number of false alarms for non-fault events) / (Total number of warnings) × 100%.

[0056] Comparison of experimental data:

[0057] During the testing period, the system triggered 217 warnings, including 42 actual fault events (valid warnings) and 175 non-fault events (false alarms). The specific performance of this invention compared to the traditional threshold method is shown in Table 2.

[0058] Table 2

[0059]

[0060] False alarm reduction rate = Traditional false alarm rate - False alarm rate of this invention × 100% = 80.6% 80.6% - 51.1% × 100% ≈ 36.6% ≈ 37%, which shows that the false alarm rate has been reduced significantly.

[0061] The reasons for the decrease in false alarm rate can be summarized as follows:

[0062] (1) The present invention reduces noise interference from single data sources through multi-source data fusion. Traditional thresholding methods rely solely on electrical parameters (such as temperature), which are susceptible to noise from environmental fluctuations (such as increased humidity due to heavy rain) or human operations (such as repairing touch sensors). The present invention integrates environmental (temperature and humidity / water immersion) and video data (human behavior) through Node-RED protocol conversion, and performs cross-validation using multi-dimensional features.

[0063] Example: When the temperature sensor triggers 78℃ (close to the threshold of 80℃), if the video AI detects that "no one is near the equipment" and the ambient humidity does not exceed the standard (<70%), it is determined to be "instantaneous sensor drift" and no false alarm is triggered.

[0064] (2) Multi-task monitoring model learns complex patterns. Traditional threshold methods set fixed thresholds based on experience, which cannot adapt to dynamic scenarios such as equipment aging and load changes. This invention learns the implicit "normal-abnormal" patterns of equipment through a multi-task AI model (safety risk prediction / anomaly detection / energy efficiency analysis).

[0065] Example: When the transformer load rate increases from 60% to 85%, the traditional method does not alarm because the temperature has not reached 80℃; however, the AI ​​model analyzes "load rate + winding temperature rise rate + ambient ventilation status" to identify the abnormal pattern of "sudden load increase leading to accelerated temperature rise" and triggers an early warning.

[0066] (3) Comprehensive analysis and optimization of early warning decision-making. Traditional threshold method triggers alarms in a "one-size-fits-all" manner. This invention calculates the final risk value through comprehensive analysis (integrating safety risk probability, anomaly score, and energy efficiency index), avoiding excessive alarms from a single dimension.

[0067] Example: A water immersion sensor in a low-voltage distribution box is activated (traditional method would directly trigger an alarm), but the AI ​​model detects that "there is no water accumulation in the video," "the ambient humidity has not continued to rise," and "there is no historical record of water immersion faults," and determines that "the sensor is falsely triggered," reducing the risk value to 0.4 (low risk), and no false alarm is triggered.

[0068] Experimental conclusion: This invention effectively solves the problem of high false alarm rate caused by the traditional threshold method due to reliance on single data, rigid fixed thresholds, and lack of pattern learning. The false alarm rate is reduced from 80.6% to 51.1%, a reduction of 37%, which significantly improves the accuracy and efficiency of power distribution station operation and maintenance.

[0069] In this invention, the terms "first," "second," and "third," etc., are used only to distinguish similar objects and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. The use of terms such as "upper," "lower," "left," "right," "front," and "rear" to indicate orientation or positional relationships is based on the orientation or positional relationships shown in the accompanying drawings and is only for the convenience of describing the invention, not to indicate or imply that the device referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation on the scope of protection of this invention. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0070] Furthermore, in the description of this application, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0071] The above are merely specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing upon the protection scope of the present invention.

Claims

1. A method for monitoring the safety of electricity consumption in a power distribution station, characterized in that: Includes the following steps: Step S1: Collect electrical parameters, environmental parameters, and video data within the substation. Using the DTU data transmission unit, upload the collected data to the control terminal via the communication protocol corresponding to the data type. Step S2: The control end uses the Node-RED data protocol conversion program to process the data uploaded using different communication protocols to obtain training data in a unified format. Step S3: Construct an artificial intelligence-based monitoring model and train the monitoring model using training data. The monitoring model includes a shared feature extractor, a safety risk prediction branch, an anomaly prediction branch, and an energy efficiency analysis branch connected to the shared feature extractor, and a fusion module that integrates and makes decisions based on the results obtained from the safety risk prediction branch, anomaly prediction branch, and energy efficiency analysis branch. Step S4: Trigger the corresponding alarm level based on the output of the fusion module.

2. The method for monitoring the power safety of a substation according to claim 1, characterized in that: In step S1, the substation is equipped with current sensors, voltage sensors and / or power sensors for collecting electrical parameters, temperature and humidity sensors and / or water immersion sensors for collecting environmental parameters, and video monitoring equipment for collecting video data. Each sensor and video monitoring equipment is connected to the DTU data transmission unit, and the communication protocol includes Modbus, MQTT, CoAP or HTTP.

3. The method for monitoring the power safety of a substation according to claim 2, characterized in that: In step S3, the shared feature extraction module adopts a multi-layer neural network structure. The system comprises a feature embedding layer, a temporal feature extraction layer, a high-level feature abstraction layer, and an output layer. Both the feature embedding layer and the temporal feature extraction layer employ a neural network structure. The training data X is processed by the feature embedding layer to obtain h1 = σ(W1·X + b1). h1 is then processed by the bidirectional LSTM of the temporal feature extraction layer to capture temporal features. After h2 is processed by the high-level feature abstraction layer, h3 = σ(W3·h2 + b3) is obtained. After h3 is normalized by the output layer, the shared feature F = LayerNorm(h3) is obtained. , W1 and W3 are the weight matrices of the feature embedding layer and the high-level feature abstraction layer, respectively, b1 and b3 are the bias vectors of the feature embedding layer and the high-level feature abstraction layer, respectively, and σ represents the ReLU activation function.

4. The method for monitoring the power safety of a substation according to claim 3, characterized in that: In step S3, the security risk prediction branch includes a risk input layer, a first temporal modeling layer, an attention mechanism layer, a security risk prediction layer, and a first output layer. The risk input layer is based on F. risk =W risk ·F+b risk Extract risk-related features F from shared features F. risk The first temporal modeling layer uses two LSTM layers for risk temporal prediction, resulting in r2 = LSTM2(r1). The attention mechanism layer introduces temporal attention weights α = softmax(W). α ·r2+b α ), to obtain r att =Σ(α⊙r2), the security risk prediction layer is based on y risk =softmax(W p ·r att +b p ) to predict the risk level, and based on p risk =sigmoid(U p ·r att +c p ) Perform risk probability calculation, and the first output layer outputs RiskOutput=(y risk ,p risk ,t), where r1=LSTM1(F risk ), LSTM1 represents the first LSTM layer, LSTM2 represents the second LSTM layer, ⊙ represents element-wise multiplication, W α b α W represents the weight matrix and bias vector of the attention mechanism layer. p b p U represents the weight matrix and bias vector for risk level prediction. p c p This represents the weight matrix and bias vector used in the risk probability calculation, where t is the timestamp for the safety risk prediction, and y represents the risk level. risk ∈ , Let K be the set of real numbers, and K be the set number of risk levels.

5. The method for monitoring the power safety of a substation according to claim 4, characterized in that: In step S3, the anomaly prediction branch includes an anomaly input layer, a convolutional feature extraction layer, a second temporal modeling layer, an anomaly scoring calculation layer, an anomaly type classification layer, and a second output layer. The anomaly input layer is based on F... anomaly =W anom ·F+b anom Extract anomaly-related features F from shared features F. anomaly The convolutional feature extraction layer uses one-dimensional convolution to capture local anomaly patterns, resulting in c2=MaxPooling(c1) and c1=Conv1D(F). anomaly kernel size =3), the second time series modeling layer uses two bidirectional LSTM layers to obtain b2=BiLSTM2(b1), b1=BiLSTM1(c2), and the anomaly scoring calculation layer is based on formula p anomaly =sigmoid(W s ·flatten(b2)+b s ) Calculate the anomaly score; the anomaly type classification layer is based on y type =softmax(W t ·b2[:,-1,:]+b t Predict the probability distribution of anomaly types, and output AnomalyOutput = (p anomaly ,y type ,t), where BiLSTM1 represents the first bidirectional LSTM layer, BiLSTM2 represents the second bidirectional LSTM layer, and W s b s Let b1 represent the weight matrix and bias vector of the anomaly scoring calculation layer, b2[:,-1,:] represent the hidden state taken from the last time step of the bidirectional LSTM layer output, and p... anomaly ∈[0,1].

6. The method for monitoring the power safety of a substation according to claim 5, characterized in that: The energy efficiency analysis branch includes an energy efficiency input layer, a self-attention mechanism layer, an energy efficiency index layer, an energy efficiency scoring layer, and a third output layer. The energy efficiency input layer is based on F... eff =W eff ·F+b eff Extracting energy efficiency-related features F eff Energy efficiency related characteristics F eff Inputting into the self-attention mechanism layer, we obtain F att =A·V, Power Factor Efficiency (PF) is calculated using the energy efficiency index layer. efficiency =MLP(F att [:,t pf ,:]) and load balancing balance =MLP(mean(F att The energy efficiency rating layer obtains E score =w1·PF efficiency +w2·Load balance +b, the third output layer outputs EfficiencyOutput=(PF efficiency Load balance E score ,t), where A=softmax(Q·Kᵀ / ), Q=W Q ·F eff K=W K ·F eff V=W V ·F eff W eff b eff This represents the weight matrix and bias vector of the energy-efficient input layer, where d is the dimension of K, MLP represents a multilayer perceptron, and t pf The key time step index for the power factor, F att [:,t pf ,:] indicates the critical sampling time t for the power factor. pf eigenvectors, mean(F) att This indicates that Fatt is averaged over time, where w1 and w2 are weighting coefficients, b is the bias term for the energy efficiency score, and W... Q W K W V These represent the transformation weight matrices for Query, Key, and Value in the self-attention mechanism, respectively.

7. The method for monitoring the power supply safety of a substation according to any one of claims 1 to 6, characterized in that: The fusion module utilizes the DS evidence theory to fuse the outputs of the safety risk prediction branch, the anomaly prediction branch, and the energy efficiency analysis branch.

8. The method for monitoring the power safety of a substation according to claim 6, characterized in that: Using loss function L tota =λ1L_ risk +λ2L_ anomaly +λ3L_ efficiency The monitoring model is trained using an alternating training strategy, where L risk =CrossEntropy(y risk ,ŷ risk )+MSE(p risk ,p̂ risk ), L anomaly =BCE(p anomaly ,p̂ anomaly )+CrossEntropy(y type ,ŷ type ), L efficiency =MSE(PF efficiency ,PF̂ efficiency )+MSE(Load balance L̂oad balance )+MSE(E score ,Ê score ), ŷ risk The true label indicating the risk level, p̂ risk The true label representing the probability of risk, p̂ anomaly For the true label of outlier scores, ŷ type For the true label of the exception type, PF̂ efficiency For the true value of power factor efficiency, L̂oad balance For the true value of load balancing, Ê score For the true value of the energy efficiency score, CrossEntropy(.) represents the cross-entropy loss function, MSE(.) represents the mean squared error loss function, BCE(.) represents the binary cross-entropy loss function, and λ1, λ2, and λ3 are the weighting coefficients of the three branch loss functions.

Images