Intelligent fault diagnosis and self-adaptive self-healing control method for station service power system of self-provided power station of hydropower station

Abstract

The invention discloses an intelligent fault diagnosis and self-adaptive self-healing control method for an auxiliary power system of a self-provided power station of a hydropower station, and belongs to the technical field of power supply system intelligence. Comprising a fault diagnosis stage and a self-healing control stage, and the intelligent fault diagnosis implementation process comprises the following steps: S1, data preprocessing; S2, multi-modal fusion fault classification; s3, fault tracing and positioning; s4, model self-optimization updating is carried out; the self-adaptive self-healing control algorithm comprises a conventional double-power-supply-mode self-healing control logic, an extreme voltage loss scene self-healing control logic, a load and power supply cooperative scheduling logic and a core load zero-interruption guarantee logic. The fault diagnosis model realizes accurate diagnosis and traceability of multi-scene faults of the dual power supply system through multi-modal fusion and mode adaptive switching, and the positioning accuracy reaches a loop level; according to the self-adaptive self-healing control algorithm, self-adaptive control over the whole process of conventional operation, fault self-healing and mode switching is achieved, and the operation stability and reliability of the auxiliary power system are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent power supply system technology, specifically to an intelligent fault diagnosis and adaptive self-healing control method for the power supply system of a self-provided power station in a hydropower station. Background Technology

[0002] The plant power system is a critical support system for the production and operation of hydropower stations, responsible for supplying power to core loads such as unit auxiliary equipment, monitoring systems, and gate control. Since hydropower station plant power systems generally adopt a dual power supply mode of "self-generated power + grid power," the operating scenarios cover various states, including conventional power supply, dual power supply parallel operation, extreme pressure loss emergency, and pressure loss recovery transition. Fault characteristics differ significantly under different scenarios (e.g., circulating current faults under dual power supply parallel operation, current surge faults under extreme pressure loss). Traditional single models (such as those relying solely on CNN or LSTM) struggle to comprehensively extract fault characteristics across multiple scenarios, resulting in low diagnostic accuracy. Furthermore, existing fault diagnosis lacks a closed-loop connection with data preprocessing, fault tracing, and model updates, failing to form a complete intelligent diagnostic process and thus failing to meet the long-term operational needs of the system.

[0003] In terms of self-healing control, existing technologies are mostly "passive response after a fault," failing to incorporate a predictive-self-healing collaborative mechanism based on dual power supply characteristics. Furthermore, they are prone to problems such as excessive voltage fluctuations and power outages to core loads during switching. In addition, they lack coordination with the hydropower station's power generation dispatch system, failing to adapt to unit output adjustments in advance, resulting in delayed power supply mode switching and impacting system stability. Therefore, there is an urgent need for a fault diagnosis method covering the entire process of "data preprocessing - fault classification - source tracing - model optimization," as well as an adaptive self-healing control method that combines dual power supply characteristics with power generation dispatch coordination. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide an intelligent fault diagnosis and adaptive self-healing control method for the power system of a self-contained power station in a hydropower station. By constructing a multimodal fusion deep learning model, it realizes accurate fault diagnosis and source tracing in multiple scenarios. Combined with the dual power supply characteristics, an adaptive self-healing control algorithm is designed to ensure the continuity and stability of power supply.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The intelligent fault diagnosis and adaptive self-healing control method for the power supply system of the self-provided power station of a hydropower station includes a fault diagnosis stage and a self-healing control stage. The fault diagnosis stage adopts a fault diagnosis model based on deep learning to realize intelligent fault diagnosis of the dual power supply system of the power station in all scenarios. The self-healing control stage is implemented by an adaptive self-healing control algorithm. The implementation process of intelligent fault diagnosis based on deep learning fault diagnosis models includes the following steps: S1. Data preprocessing: Collect full-scenario data of the plant power system under the conditions of "normal power supply + extreme voltage loss". After labeling the collected data with scenario tags, use conditional generative adversarial network CGAN to supplement scarce fault samples. Then, use wavelet transform to filter load disturbance noise in order to extract pure fault features. S2. Multimodal fusion fault classification: The preprocessed fault features are input into the CNN-LSTM-Transformer hybrid model. The CNN extracts voltage and current waveform features, the LSTM captures temporal dependencies, and the Transformer attention mechanism focuses on key features of dual power supply, and outputs the fault classification results. S3. Fault source tracing and location: Construct a topological relationship graph of "power supply head - bus tie cabinet - feeder cabinet - load" based on graph neural network (GNN), and use the signal transmission relationship between nodes to trace the source of the fault in reverse. S4. Model self-optimization and update: Regularly collect new dual power supply mode data and equipment calibration data, and fine-tune model parameters using gradient descent method; The adaptive self-healing control algorithm includes self-healing control logic for conventional dual power supply mode, self-healing control logic for extreme undervoltage scenarios, load and power supply coordinated scheduling logic, and core load zero-interruption guarantee logic.

[0006] The processing steps of the data preprocessing module in S1 mentioned above specifically include: S11. Collect dual power supply mode data and environmental and load correlation data. Dual power supply mode data includes bus voltage, current fluctuation, power flow data and plant load switching status and power load change data under normal power supply scenarios, and voltage surge / dip waveforms, current surge peaks, switch action timing data and switching transition data after voltage recovery under extreme power loss scenarios. Correlation data includes hydropower station power generation conditions, seasonal climate data, and equipment operating parameter data. Hydropower station power generation conditions include the number of operating units and output data, seasonal climate data includes temperature, humidity and precipitation, and equipment operating parameter data includes bus tie switch circuit breaker temperature rise and transformer insulation resistance. S12. The collected data are labeled with "Grid Power Supply Status," "Self-Generated Power Supply Status," "Dual Power Supply Parallel Status," "Full Station Power Loss Emergency Status," and "Power Loss Recovery Transition Status." Scarce fault samples include voltage and current time-series data corresponding to dual-power switching faults and spontaneous power fluctuation faults. When using a Conditional Generative Adversarial Network (CGAN) to generate simulated fault data to supplement training samples, the voltage amplitude range and current change rate of the real data are used as constraints to ensure that the difference between the simulated data and the real data distribution is ≤5%. Load disturbance noise is filtered through wavelet transform to extract pure fault features. The decomposition and reconstruction formulas for wavelet transform are as follows: Decomposition formula: ; ; ; Reconstructing the formula: ; in, It is a time series. The original signal, The number of layers in the decomposition. , For wavelet decomposition filters in the time domain, , For wavelet reconstruction filters in the time domain, For signal In the Wavelet coefficients in the low-frequency part of the layer, For signal In the Wavelet coefficients of the high-frequency part of the layer.

[0007] The multimodal fusion diagnostic model in S2 above is a CNN-LSTM-Transformer hybrid model, and the formula is as follows: ; in, The voltage and current data are after preprocessing. This indicates that waveform features of voltage and current are extracted using CNN. This indicates that LSTM is used to capture temporal dependencies. This is the weight matrix for the Transformer attention mechanism, used to focus on key features in the dual-power mode. The results of the fault classification; Model training uses the cross-entropy loss function as the objective function, and the formula is: ; in, For the sample size, Number of fault categories For the first The sample belongs to the first The true label (0 or 1) for the type of fault. For the model to predict the first The sample belongs to the first To determine the probability of each type of fault, the Adam optimizer was used to minimize the loss function during training. The learning rate was set to 0.001~0.005, and the number of iterations was 300~800 to ensure that the model's classification accuracy was ≥99.5%. The model also includes a mode adaptive switching module, which receives the power supply mode signal of the bus tie cabinet in real time and automatically adjusts the feature extraction weights. During normal power supply, it focuses on voltage harmonics and load fluctuation characteristics, setting the weight coefficients of voltage harmonics and load fluctuation characteristics to 0.6~0.8. During extreme voltage loss, it focuses on current surge and standby equipment status characteristics, setting the weight coefficients of current surge and standby equipment status characteristics to 0.7~0.9, so as to avoid diagnostic bias caused by mode confusion.

[0008] The formula for the fault tracing model in S3 above is: ; in, For nodes The output characteristics, , They are nodes and its neighboring nodes Input features, For nodes The set of neighboring nodes, , This is the weight matrix. As the activation function, the fault source is traced back through the signal transmission relationship between nodes, and the positioning accuracy reaches the loop level.

[0009] The parameter update formula for the model self-optimization module in S4 above is: ; in, Let be the model parameters at time t. For the updated model parameters, For learning rate, For loss function, For newly acquired dual-power supply mode data and equipment calibration data, the model parameters are fine-tuned regularly to ensure long-term stability of diagnostic accuracy. The model self-optimization update cycle is once a month. The equipment calibration data includes circuit breaker operating characteristic data (opening and closing time, operating voltage) and transformer insulation parameters (insulation resistance, dielectric loss) after maintenance during the dry season. The sample size of newly acquired data is no less than 10% of the historical training sample size for each update.

[0010] The self-healing control logic in the conventional dual-power supply mode described above includes dual-power supply parallel control and adaptive switching of power supply modes. The circulating current control formula for the dual-power supply parallel control is as follows: ; in, For circulating current when two power supplies are connected in parallel, The rated current of the bus tie cabinet is used to achieve circulating current control by adjusting the closing phase of the circuit breaker and the current limiting resistor of the bus tie cabinet. The adaptive switching logic for power supply mode is as follows: when the voltage fluctuation of self-generated power is ≤±2% and the load meets the needs of the plant, self-generated power supply is automatically prioritized; when the voltage of self-generated power is <360V, it switches to grid power supply within 0.5 seconds; when the grid fails, self-generated power supply is activated at full load, and if self-generated power is insufficient, non-core loads are automatically cut off.

[0011] The self-healing control logic under the extreme pressure loss scenario mentioned above includes rapid pressure loss response and pressure loss recovery transition control. The rapid pressure loss response completes the actions of disconnecting all non-core load circuits, starting the backup transformer, closing the backup power supply incoming switch, and restoring power supply to the core load within a set time. The undervoltage recovery transition control adopts a "step transition" strategy. First, the backup power supply is used to power the core load and the stability of the recovery power supply is detected to reach the set time. Then, voltage phase synchronization technology is used to switch the core load to the recovery power supply with a phase error of ≤0.5°. Finally, the non-core loads are restored in batches.

[0012] The aforementioned load and power supply coordinated scheduling logic includes dynamic load priority management and power balance control. The power balance control formula is as follows: ; in, This refers to the power supply capacity, which is the sum of self-generated power capacity and grid power supply capacity. For core load power, For the first Non-core load power, This is the non-core load allocation factor, which is 0 or 1; when the load exceeds the power supply capacity, it will automatically switch to offload. Set to 0 to cut off low-priority non-core loads; when there is excess power supply capacity, start the energy storage device to store energy.

[0013] The aforementioned core load zero-interruption protection logic adopts "uninterrupted power supply transition" technology. Through the soft-start control of the bidirectional thyristor in the bus tie cabinet and the backup power supply, the voltage fluctuation control formula is as follows: ; in, This refers to the instantaneous voltage during the switching process. The rated voltage ensures that the core load does not restart or crash.

[0014] The aforementioned adaptive self-healing control algorithm communicates in real time with the hydropower station's power generation dispatch system to obtain unit operation plans and output adjustment instructions, and adjusts the power supply strategy in advance. The formula is as follows: ; in, For control strategies, For power generation dispatch instructions, It is in dual power supply mode. It enables advance adaptation to load status, power supply mode switching, and load scheduling.

[0015] The present invention discloses an intelligent fault diagnosis and adaptive self-healing control method for the auxiliary power system of a hydropower station. The fault diagnosis model achieves accurate diagnosis and source tracing of faults in multiple scenarios of the dual power supply system through multi-modal fusion and adaptive mode switching, with a diagnosis accuracy of ≥99.5% and a positioning accuracy at the loop level. The adaptive self-healing control algorithm combines the characteristics of dual power supply with power generation scheduling coordination to achieve adaptive control of the entire process of normal operation, fault self-healing, and mode switching. The core load power supply is uninterrupted, and the voltage fluctuation is ≤±1%, effectively improving the operational stability and reliability of the auxiliary power system. Attached Figure Description

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a schematic diagram of the power supply system of the self-provided power station of the hydropower station according to the present invention; Figure 2 This is a schematic diagram of the primary power supply system for the plant. Figure 3 This is a schematic diagram of the fault diagnosis process based on deep learning in this invention; Figure 4 This is a logic diagram of the self-healing control method of the present invention.

[0017] In the diagram: 1. Hydropower unit; 2. Step-up transformer; 3. Power grid; 4. Hydropower plant area; 5. Plant auxiliary transformer; 6. Plant auxiliary power supply system; 7. Energy storage device; 8. Backup power supply. Detailed Implementation

[0018] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0019] like Figure 1 The diagram shown is a schematic of the power system for a hydropower station's self-contained power plant, including: 1. Hydropower generating unit; 2. Step-up transformer; 3. Power grid; 4. Hydropower plant area; 5. Power plant auxiliary transformer; 6. Power plant auxiliary power supply system; 7. Energy storage device; and 8. Backup power supply. Figure 2 The diagram shown is a primary wiring diagram of the plant's power system, which includes components such as busbars, reactive power compensation, metering, power receiving, and power feeding.

[0020] A method for intelligent fault diagnosis and adaptive self-healing control of the power supply system of a self-provided power station in a hydropower station includes a fault diagnosis stage and a self-healing control stage. The fault diagnosis stage is implemented using a fault diagnosis model based on deep learning, and the self-healing control stage is implemented using an adaptive self-healing control algorithm.

[0021] I. Fault Diagnosis Model Based on Deep Learning This model, centered on "full-scenario data support - multi-modal feature fusion - precise source tracing - dynamic optimization," forms a complete intelligent diagnostic process. The logical connections between each step are as follows: (a) Data Preprocessing Module 1. Full-scene data collection To address the dual power supply characteristics of hydropower stations, electrical parameters and related data were collected under the scenario of "conventional power supply + extreme power outage": In a typical power supply scenario: the bus tie cabinet receives self-generated power (the power branch of the generating unit), the bus voltage (380V±5%) when the grid is connected, three-phase current, and active / reactive power data, and simultaneously records the switching status and load changes (10kW~500kW) of the plant's water pumps, lighting, maintenance power supplies, and other loads. Extreme voltage loss scenario: When the entire station loses voltage, the voltage surge (380V→420V) / drop (380V→0V) waveform, current surge peak (≤2000A), circuit breaker action sequence (opening and closing time ≤0.1 seconds) data of the standby transformer (200kVA) during startup, and voltage transient (lasting 0.5~2 seconds) and current fluctuation data during the switching transition after the voltage loss is restored; Related data: Number of operating hydropower station units (1~4 units), power output (500kW~2000kW), seasonal temperature and humidity (-20℃~40℃, humidity 30%~90%), temperature rise of bus tie switch circuit breaker (≤60K), and insulation resistance of transformer (≥1000MΩ).

[0022] 2. Data Augmentation and Denoising Scarce Sample Enhancement: To address the scarcity of samples, such as dual-power switching faults (circulating current exceeding 100A during parallel operation) and spontaneous power fluctuation faults (voltage drop exceeding 20V), CGAN is used to generate simulated data. Constrained by the voltage amplitude range (360V~400V) and current change rate (≤50A / ms) of real data, the distribution difference between the generated samples and real samples is ≤5%. After supplementation, the proportion of scarce samples increases from 5% to 30%. Wavelet denoising: The db4 wavelet basis function is used, and the number of decomposition layers is set to 3. Load disturbance noise such as current surge (50~100A) caused by the start and stop of the water pump in the plant is filtered out. Pure fault features such as current surge during short circuit (>200A / ms) and zero-sequence current abnormality during grounding (>5A) are extracted. After denoising, the signal-to-noise ratio of the signal is improved to more than 30dB.

[0023] The decomposition and reconstruction formulas for wavelet transform are as follows: Decomposition formula: ; ; ; Reconstructing the formula: ; in, It is a time series. The original signal, The number of layers in the decomposition. , For wavelet decomposition filters in the time domain, , For wavelet reconstruction filters in the time domain, For signal In the Wavelet coefficients in the low-frequency part of the layer, For signal In the Wavelet coefficients of the high-frequency part of the layer.

[0024] (II) Multimodal fusion fault classification 1. “CNN-LSTM-Transformer” hybrid model structure A hybrid "CNN-LSTM-Transformer" model is constructed. CNN extracts voltage and current waveform features, LSTM captures temporal dependencies, and the Transformer attention mechanism focuses on key features of dual power supply. CNN module: It uses 3 convolutional layers with kernel sizes of 3×1, 5×1 and 7×1, and output channels of 16, 32 and 64 respectively. It extracts local features of voltage and current waveforms, such as short-circuit spikes and harmonic distortion. LSTM module: Set up 2 hidden layers, with 128 neurons in each layer, to capture fault timing dependencies (such as voltage dips lasting 1~3 seconds, current surges lasting 0.1~0.5 seconds). Transformer module: Set up 8 attention heads, weight matrix The dimensions are 256×256, focusing on key characteristics of dual power supply, such as phase difference > 5° when connected in parallel and timing deviation > 0.05 seconds when switching.

[0025] The model formula is: ; in, The voltage and current data are after preprocessing. This indicates that waveform features of voltage and current are extracted using CNN. This indicates that LSTM is used to capture temporal dependencies. This is the weight matrix for the Transformer attention mechanism, used to focus on key features in the dual-power mode. The fault classification results show an accuracy rate of ≥99.5%.

[0026] The model includes a mode adaptive switching module that automatically adjusts the feature extraction weights under different power supply modes to avoid diagnostic bias caused by mode confusion.

[0027] 2. Model Training and Optimization Objective function: The cross-entropy loss function is used to quantify the difference between the model's predicted probability and the true label, as shown in claim 1; Training parameters: The Adam optimizer was used, the learning rate was set to 0.003, the number of iterations was 500, the batch size was set to 32, and the training set and test set were divided in a 7:3 ratio. Adaptive mode switching: The model receives the power supply mode signal of the bus tie cabinet in real time. During normal power supply, the weight of voltage harmonics and load fluctuation characteristics is set to 0.7, and during extreme voltage loss, the weight of current surge and standby equipment status characteristics is set to 0.8, ensuring that the diagnostic accuracy is ≥99.5% in different scenarios.

[0028] (III) Fault source tracing and location 1. Construction of GNN topological association graph A topology graph of "power supply head - bus tie cabinet - feeder cabinet - load" is constructed based on GNN. The source of the fault is traced back through the signal transmission between nodes. The nodes include: power supply head node (self-generated power side, grid side), bus tie cabinet node, feeder cabinet node (1~10), and load node (core load: monitoring system, gate control; non-core load: office lighting, redundant water pump). The edge weight is set according to the electrical connection strength between the equipment. For example, the edge weight between the bus tie cabinet and the feeder cabinet is set to 0.9 (direct power supply), and the edge weight between the feeder cabinet and the non-core load is set to 0.6 (indirect power supply).

[0029] The model formula is: ; in, For nodes The output characteristics, , They are nodes and its neighboring nodes Input features, For nodes The set of neighboring nodes, , This is the weight matrix. As the activation function, the fault source is traced back through the signal transmission relationship between nodes, and the positioning accuracy reaches the loop level.

[0030] 2. Traceability Process When an abnormal bus voltage (<360V) is detected, the fault source is traced in reverse through the GNN node characteristics: abnormal voltage characteristics of bus tie cabinet node → abnormal current characteristics of feeder cabinet node → abnormal power characteristics of load node (such as voltage dip caused by lightning strike on the grid side line, or frequency drift caused by abnormal speed controller of self-generated power unit). The positioning accuracy reaches the loop level (such as a fault in the 3 feeder cabinet loop), and the tracing time is ≤0.5 seconds.

[0031] (iv) Model self-optimization update 1. Update cycle and data sources Collect new dual power supply mode data (≥1000 records) every month, and import equipment calibration data after maintenance during the dry season (circuit breaker opening and closing time 0.08~0.12 seconds, transformer insulation resistance 1000~2000MΩ).

[0032] The model parameters are fine-tuned periodically using newly collected data. The model parameter update formula is as follows: ; in, Let be the model parameters at time t. For the updated model parameters, For learning rate, For loss function, For newly acquired dual-power-mode data and equipment calibration data, the model parameters are fine-tuned periodically to ensure long-term stability of diagnostic accuracy.

[0033] 2. Parameter update mechanism Mini-batch gradient descent was used with a learning rate of 0.0005. Only the weights of the CNN convolutional kernel, the weights of the LSTM hidden layer, and the weights of the Transformer attention head were fine-tuned during each update. After the update, the diagnostic accuracy of the model on the new data remained ≥99.5%, thus avoiding model overfitting.

[0034] II. Adaptive Self-Healing Control Algorithm The algorithm aims for "stable operation under normal conditions, self-healing from extreme faults, load-coordinated scheduling, and zero core interruptions," combining dual power supply characteristics with power generation scheduling coordination to achieve full-process adaptive control. (I) Self-healing control logic in conventional dual power supply mode 1. Dual power supply parallel control By adjusting the closing phase (phase difference ≤ 2°) and the current-limiting resistor (resistance value 10~30Ω) of the bus tie switch (NORAD Ex9A25H), the circulating current is controlled to be ≤ 5% of the rated current (1000A) (i.e. ≤ 50A). If the circulating current is detected to be > 50A, the bus tie switch is immediately disconnected and switched to single power supply to avoid overheating of the bus (temperature rise ≤ 60K).

[0035] 2. Adaptive switching of power supply mode Self-generated power priority: When the voltage fluctuation of self-generated power is ≤±2% (372V~388V) and the load meets the needs of the plant (self-generated power capacity > total load power 10%), disconnect the power supply on the grid side to reduce the cost of grid electricity, and the switching time is ≤0.3 seconds; Grid standby switching: When the self-generated power voltage is <360V (sudden drop in unit output), switch to grid power supply within 0.5 seconds, and at the same time adjust the reactive power compensation device (Chint NXU series) of the feeder cabinet to stabilize the bus voltage at 380V±2%; Power outage response: When the power grid fails, the self-generated power will be activated immediately to supply power at full load. If the self-generated power capacity is insufficient (< total load power), non-core loads (such as office lighting and redundant water pumps) will be automatically cut off to ensure power supply to core loads.

[0036] (II) Self-healing control logic for extreme pressure loss scenarios 1. Rapid response to pressure loss Upon receiving the "station-wide power failure" diagnostic signal, complete the following within 1 second: Disconnect all non-core load circuits (by tripping the molded case circuit breaker in the power supply cabinet). Start the backup transformer and simultaneously activate the preheating device when the temperature is low in winter (<0℃) to preheat to above 5℃; Close the backup power supply switch to restore power to the core load; the power interruption time should be ≤1 second.

[0037] 2. Transition control for pressure loss recovery Stability test: With the backup power supply as the core load, the voltage (380V±2%) and frequency (50Hz±0.1Hz) of the recovery power supply (grid / self-generated power) are continuously tested for 10 seconds. After confirming stability, the switching process is initiated. Phase synchronization switching: Voltage phase synchronization technology is adopted to control the phase error during switching to ≤0.5° and avoid current surge (≤100A). Load restoration in batches: Non-core loads are restored in the order of "low power (lighting 10kW) → medium power (maintenance power 50kW) → high power (water pump 200kW)", with a 5-second interval between each batch. Finally, the backup power supply is disconnected and the power supply is restored to normal.

[0038] (III) Load and power supply coordinated scheduling logic 1. Dynamic load priority management During power generation: Priority will be given to ensuring power supply to the unit's auxiliary equipment (50kW lubricating oil pump, 100kW cooling system), with a priority of 1; During the flood discharge period: priority will be given to ensuring power supply for gate control (80kW) and hydrological monitoring (20kW) equipment, with a priority level of 1; During maintenance: Priority will be given to ensuring power supply for maintenance (50kW) and safety lighting (5kW), with a priority of 1; Non-core loads are prioritized to 2-3, and during self-healing, they are removed from low to high priority.

[0039] 2. Power balance control Based on the balance between power supply capacity and load power, the system automatically adjusts the load on / off status, dynamically manages load priorities, and automatically disconnects low-priority loads or activates energy storage devices. The power balance control formula is as follows: ; in, For power supply capacity, For core load power, For the first Non-core load power, This is the non-core load allocation factor (0 or 1). When the load exceeds the power supply capacity, the load will be automatically reduced. Set to 0 to cut off low-priority non-core loads; when there is excess power supply capacity, start the energy storage device to store energy.

[0040] When the power supply capacity (800kW self-generated + 500kW grid = 1300kW) < total load power (300kW core + 1200kW non-core = 1500kW), the low-priority non-core load (300kW) is cut off, so that the total load power is reduced to 1200kW, thus satisfying the power balance. When the power supply capacity (1000kW self-generated + 500kW grid power = 1500kW) > the total load power (1200kW) and the excess power > 50kW, the backup battery energy storage (300kW energy storage capacity) is activated to avoid energy waste.

[0041] (iv) Core load zero-interruption guarantee logic and coordination with power generation dispatch 1. Zero-interrupt technology The system employs a bidirectional thyristor (withstanding voltage 1200V) in the bus tie cabinet and a soft-start control for the backup power supply. During switching, the instantaneous voltage fluctuation is ≤±1% (376.2V~383.8V), ensuring that the core loads such as the monitoring system and PLC controller do not restart or crash, thus meeting the requirements for safe operation of the hydropower station.

[0042] The "uninterruptible power supply transition" technology is adopted to control voltage fluctuations to ≤±1% during the switching process, ensuring that the core load does not restart or crash. The voltage fluctuation control formula is as follows: ; in, This refers to the instantaneous voltage during the switching process. The rated voltage ensures that the core load does not restart or crash.

[0043] 2. Coordination with power generation dispatch The algorithm communicates in real time with the hydropower station's power generation dispatch system to adjust the power supply strategy in advance, adapting to the unit's operation plan and output adjustment. The coordination formula is as follows: ; in, For control strategies, For power generation dispatch instructions, It is in dual power supply mode. It enables advance adaptation to load status, power supply mode switching, and load scheduling.

[0044] 2. Coordination with power generation dispatch The algorithm communicates with the hydropower station's power generation dispatch system via Ethernet to obtain the unit operation plan (e.g., adding one unit the next day, increasing output by 500kW), and adjusts the power supply strategy one hour in advance. Increase the preparation for receiving spontaneous power and adjust the voltage regulation module (thyristor voltage regulation circuit) of the bus tie cabinet to adapt to higher voltage (390V±2%). If a shutdown for maintenance is planned, switch to grid power 30 minutes in advance to avoid power outages in the plant due to the shutdown.

[0045] Example 1: Training and Application of Fault Diagnosis Model 1. Data Acquisition and Preprocessing Data on bus voltage (380V±5%), current fluctuations, and power flow under normal power supply scenarios in the 0.4kV auxiliary power system of a hydropower station were collected. Data on voltage surge / dip waveforms and current surge peaks during backup power startup under extreme undervoltage scenarios were also collected. A three-dimensional database was constructed by combining this data with hydropower station power generation conditions and seasonal climate data. Scenario labels were added to the data, and CGAN was used to generate simulated data such as dual-power switching faults and spontaneous power fluctuation faults. Load disturbance noise was filtered through wavelet transform, and the wavelet decomposition layer was set to 3 levels to extract pure fault features.

[0046] 2. Model Training A hybrid CNN-LSTM-Transformer model was constructed. The CNN used three convolutional layers with kernel sizes of 3×1, 5×1, and 7×1. The LSTM had two hidden layers with 128 neurons per layer. The Transformer had eight attention heads. The labeled dataset was divided into training and test sets in a 7:3 ratio. During training, the Adam optimizer was used with a learning rate of 0.001 and 500 iterations. The model achieved a classification accuracy of 99.7%.

[0047] 3. Fault Diagnosis and Source Tracing The model receives power supply mode signals from the bus tie cabinet in real time. During normal power supply, it focuses on extracting voltage harmonics and load fluctuation features, while during extreme power outages, it focuses on extracting current surges and standby equipment status features. When a sudden change in bus current is detected, the model traces back through the GNN topology graph to accurately locate the source of the fault as either a lightning strike on the grid-side line or damage to the filter capacitor in the local feeder cabinet.

[0048] Example 2: Application of Adaptive Self-Healing Control Algorithm 1. Normal dual power supply mode operation When the self-generated power voltage fluctuation is ±1.5% and the load meets the plant's requirements, the algorithm automatically switches to self-generated power supply and disconnects the grid power supply. When the unit output drops sharply, causing the self-generated power voltage to drop to 350V, the algorithm switches to grid power supply within 0.4 seconds and simultaneously adjusts the reactive power compensation device to stabilize the bus voltage to 380V.

[0049] 2. Self-healing in extreme decompression scenarios Upon receiving the "station-wide power failure" diagnostic signal, the algorithm disconnects all non-core load circuits, starts the backup transformer, and closes the backup power supply incoming switch within 0.8 seconds to restore power to the core loads. After the grid restores power, the stability of the restored power supply is checked for 10 seconds. Once the voltage and frequency are confirmed to be stable, voltage phase synchronization technology (phase error 0.3°) is used to switch the core loads to grid power, and then the non-core loads are restored in batches.

[0050] 3. Load and power supply coordinated scheduling When the power supply capacity is 800kW, the core load power is 500kW, and the total power of the non-core load is 400kW, the algorithm automatically cuts off low-priority non-core loads (power 100kW) to ensure power balance; when the self-generated power is sufficient and the power supply capacity is 200kW in excess, the energy storage device is activated to store energy.

[0051] 4. Zero-interruption guarantee for core load During the dual power supply switching process, the instantaneous voltage is 378V and the voltage fluctuation is -0.5% through the bidirectional thyristor in the bus tie cabinet and the soft start control of the backup power supply, which meets the requirements of the core load (monitoring system, PLC controller) to not restart or crash.

Claims

1. A method for intelligent fault diagnosis and adaptive self-healing control of the power supply system of a self-contained hydropower station, characterized in that, It includes a fault diagnosis stage and a self-healing control stage. The fault diagnosis stage uses a deep learning-based fault diagnosis model to realize intelligent fault diagnosis of the dual power supply system for plant and auxiliary power in all scenarios. The self-healing control stage is implemented using an adaptive self-healing control algorithm. The implementation process of intelligent fault diagnosis based on deep learning fault diagnosis models includes the following steps: S1. Data preprocessing: Collect full-scenario data of the plant power system under "normal power supply + extreme voltage loss". After labeling the collected data with scenario tags, use conditional generative adversarial network CGAN to supplement scarce fault samples. Then, use wavelet transform to filter load disturbance noise to extract pure fault features. S2. Multimodal fusion fault classification: The preprocessed fault features are input into the CNN-LSTM-Transformer hybrid model. The CNN extracts voltage and current waveform features, the LSTM captures temporal dependencies, and the Transformer attention mechanism focuses on key features of dual power supply, and outputs the fault classification results. S3. Fault source tracing and location: Construct a topological relationship graph of "power supply head - bus tie cabinet - feeder cabinet - load" based on graph neural network (GNN), and use the signal transmission relationship between nodes to trace the source of the fault in reverse. S4. Model self-optimization and update: Regularly collect new dual power supply mode data and equipment calibration data, and fine-tune model parameters using gradient descent method; The adaptive self-healing control algorithm includes self-healing control logic for conventional dual power supply mode, self-healing control logic for extreme undervoltage scenarios, load and power supply coordinated scheduling logic, and core load zero-interruption guarantee logic.

2. The intelligent fault diagnosis and adaptive self-healing control method for the power supply system of a self-contained hydropower station according to claim 1, characterized in that, The processing steps of the data preprocessing module in S1 specifically include: S11. Collect dual power supply mode data and environmental and load correlation data. Dual power supply mode data includes bus voltage, current fluctuation, power flow data and plant load switching status and power load change data under normal power supply scenarios, and voltage surge / dip waveforms, current surge peaks, switching action timing data and switching transition data after voltage loss recovery under extreme power loss scenarios. Correlation data includes hydropower station power generation conditions, seasonal climate data, and equipment operating parameter data. S12. The collected data are labeled with "Grid Power Supply Status," "Self-Generated Power Supply Status," "Dual Power Supply Parallel Status," "Full Station Power Loss Emergency Status," and "Power Loss Recovery Transition Status." Scarce fault samples include voltage and current time-series data corresponding to dual-power supply switching faults and spontaneous power fluctuation faults. When using a conditional generative adversarial network (CGAN) to generate simulated fault data to supplement training samples, the voltage amplitude range and current change rate of the real data are used as constraints. Load disturbance noise is filtered through wavelet transform to extract pure fault features. The decomposition and reconstruction formulas for wavelet transform are as follows: Decomposition formula: ； ； ； Reconstructing the formula: ； in, It is a time series. The original signal, The number of layers in the decomposition. , For wavelet decomposition filters in the time domain, , For wavelet reconstruction filters in the time domain, For signal In the Wavelet coefficients in the low-frequency part of the layer, For signal In the Wavelet coefficients of the high-frequency part of the layer.

3. The intelligent fault diagnosis and adaptive self-healing control method for the self-provided power station auxiliary power system of a hydropower station according to claim 1, characterized in that, The multimodal fusion diagnostic model in S2 is a CNN-LSTM-Transformer hybrid model, and the formula is as follows: ； in, The voltage and current data are after preprocessing. This indicates that waveform features of voltage and current are extracted using CNN. This indicates that LSTM is used to capture temporal dependencies. This is the weight matrix for the Transformer attention mechanism, used to focus on key features in the dual-power mode. The results of the fault classification; Model training uses the cross-entropy loss function as the objective function, and the formula is: ； in, For the sample size, Number of fault categories For the first The sample belongs to the first The true label (0 or 1) for the type of fault. For the model to predict the first The sample belongs to the first The probability of a type of failure; The model also includes a mode adaptive switching module, which receives the power supply mode signal of the bus tie cabinet in real time, automatically adjusts the feature extraction weights, focuses on voltage harmonics and load fluctuation characteristics during normal power supply, and focuses on current surge and standby equipment status characteristics during extreme voltage loss.

4. The intelligent fault diagnosis and adaptive self-healing control method for the power supply system of a self-contained hydropower station according to claim 1, characterized in that, The formula for the fault tracing model in S3 is as follows: ； in, For nodes The output characteristics, , They are nodes and its neighboring nodes Input features, For nodes The set of neighboring nodes, , This is the weight matrix. As the activation function, the fault source is traced back through the signal transmission relationship between nodes, and the positioning accuracy reaches the loop level.

5. The intelligent fault diagnosis and adaptive self-healing control method for the power supply system of a self-contained hydropower station according to claim 1, characterized in that, The parameter update formula for the model self-optimization module in S4 is as follows: ； in, Let be the model parameters at time t. For the updated model parameters, For learning rate, For loss function, For newly acquired dual-power-mode data and equipment calibration data, the model parameters are fine-tuned periodically to ensure long-term stability of diagnostic accuracy.

6. The intelligent fault diagnosis and adaptive self-healing control method for the self-provided power station auxiliary power system of a hydropower station according to claim 1, characterized in that, The self-healing control logic under the conventional dual power supply mode includes dual power supply parallel control and adaptive switching of power supply mode, wherein the circulating current control formula for dual power supply parallel control is: ； in, For circulating current when two power supplies are connected in parallel, The rated current of the bus tie cabinet is used to achieve circulating current control by adjusting the closing phase of the circuit breaker and the current limiting resistor of the bus tie cabinet. The adaptive switching logic for power supply mode is as follows: when the voltage fluctuation of self-generated power is ≤±2% and the load meets the needs of the plant, self-generated power supply is automatically prioritized; when the voltage of self-generated power is <360V, it switches to grid power supply within 0.5 seconds; when the grid fails, self-generated power supply is activated at full load, and if self-generated power is insufficient, non-core loads are automatically cut off.

7. The intelligent fault diagnosis and adaptive self-healing control method for the power supply system of a self-contained hydropower station according to claim 1, characterized in that, The self-healing control logic under extreme pressure loss scenarios includes rapid pressure loss response and pressure loss recovery transition control. The rapid pressure loss response completes the actions of disconnecting all non-core load circuits, starting the backup transformer, closing the backup power supply incoming switch, and restoring power supply to the core load within a set time. The undervoltage recovery transition control adopts a "step transition" strategy. First, the backup power supply is used to power the core load and the stability of the recovery power supply is detected to reach the set time. Then, voltage phase synchronization technology is used to switch the core load to the recovery power supply with a phase error of ≤0.5°. Finally, the non-core loads are restored in batches.

8. The intelligent fault diagnosis and adaptive self-healing control method for the power supply system of a self-contained hydropower station according to claim 1, characterized in that, The load and power supply coordinated scheduling logic includes dynamic load priority management and power balance control. The power balance control formula is as follows: ； in, This refers to the power supply capacity, which is the sum of self-generated power capacity and grid power supply capacity. For core load power, For the first Non-core load power, This is the non-core load allocation factor, which is 0 or 1; when the load exceeds the power supply capacity, it will automatically switch to offload. Set to 0 to cut off low-priority non-core loads; when there is excess power supply capacity, start the energy storage device to store energy.

9. The intelligent fault diagnosis and adaptive self-healing control method for the power supply system of a self-contained hydropower station according to claim 1, characterized in that, The core load zero-interruption protection logic adopts "uninterrupted power supply transition" technology. Through the soft-start control of the bidirectional thyristor in the bus tie cabinet and the backup power supply, the voltage fluctuation control formula is as follows: ； in, This refers to the instantaneous voltage during the switching process. The rated voltage ensures that the core load does not restart or crash.

10. The intelligent fault diagnosis and adaptive self-healing control method for the self-provided power station auxiliary power system of a hydropower station according to claim 1, characterized in that, The adaptive self-healing control algorithm communicates in real time with the hydropower station's power generation dispatch system to obtain unit operation plans and output adjustment instructions, and adjusts the power supply strategy in advance. The formula is as follows: ； in, For control strategies, For power generation dispatch instructions, It is in dual power supply mode. It enables advance adaptation to load status, power supply mode switching, and load scheduling.

Images