Microgrid fault self-healing control method, system and device

Abstract

The application discloses a micro-grid fault self-healing control method, system and equipment, and belongs to the field of electric power systems and automation, and comprises the following steps: acquiring the communication state between the main micro-grid and the discrete micro-grid; if the communication state is connected, acquiring a global control strategy according to the fault category; if the communication state is disconnected, constructing a long-time sequence power prediction model based on the maintenance cycle, the LSTM and the residual connection, training the model through the historical photovoltaic power generation sequence and the historical meteorological sequence, acquiring the total predicted power generation of the photovoltaic power generation in the maintenance cycle through the long-time sequence power prediction model, and acquiring the global control strategy based on the total predicted power generation and the load priority; and regulating and controlling the power consumption of the discrete micro-grid according to the global control strategy, and continuously detecting the fault recovery state, and when the fault recovery is detected, performing smooth grid connection according to the electrical parameters of the main micro-grid. The technical problem that the prior art cannot effectively regulate and control the discrete micro-grid under the working condition that the power supply and the communication are long-term disconnected is solved.

Description

Technical Field

[0001] This invention relates to the field of power systems and automation, specifically to a method, system, and equipment for self-healing control of microgrid faults. Background Technology

[0002] In discrete microgrids, fault self-healing under scenarios of distribution network failure and communication link interruption is a key technology for ensuring the continuous and reliable power supply to critical loads within the discrete microgrid. Conventional microgrid self-healing control systems rely on a stable communication link between the main microgrid and the discrete microgrid to complete data exchange and command issuance. However, in island power supply, submarine cables serve dual functions of power transmission and fiber optic communication. When a cable breaks, the island microgrid simultaneously loses its external main grid power supply channel and its communication channel with the upper-level master station, resulting in a dual failure of power supply and communication. Moreover, the repair cycle for such faults is usually long. Current mainstream microgrid control schemes rely on centralized calculations by an upper-level central controller to achieve global scheduling decisions. Once the communication link is broken, local devices cannot independently complete scheduling logic calculations, and the discrete microgrid will lose its autonomous intelligent self-healing regulation capability, making it difficult to maintain stable power supply to loads within the discrete microgrid. Therefore, how to effectively regulate discrete microgrids under conditions of long-term power and communication disconnection, thereby achieving autonomous fault self-healing regulation of discrete microgrids, is a technical challenge that current technologies cannot solve. Summary of the Invention

[0003] To address the technical challenge of effectively controlling discrete microgrids under conditions of prolonged power and communication outages, this invention provides a microgrid fault self-healing control method, system, and equipment. When communication is disconnected, a long-term power prediction model is constructed using maintenance cycles, LSTM (Laser-Based Memory) technology, and residual connections. This overcomes the insufficient accuracy of long-term power output prediction using a single LSTM. The total predicted photovoltaic power generation within the maintenance cycle is obtained based on the long-term power prediction model, and a global control strategy is derived by combining load priority. This solves the technical problem of effectively controlling discrete microgrids under conditions of prolonged power and communication outages.

[0004] To address the aforementioned technical problems, this invention provides a microgrid fault self-healing control method, comprising the following steps: The communication status between the main microgrid and the discrete microgrids is continuously monitored through a heartbeat mechanism; If the communication status is connected, a global control strategy is obtained based on the fault category. If the communication status is disconnected, a long-term power prediction model is constructed based on the maintenance cycle through LSTM and residual connections. The long-term power prediction model is trained through historical photovoltaic power generation sequences and historical meteorological sequences. The total predicted power generation of photovoltaics within the maintenance cycle is obtained through the long-term power prediction model. A global control strategy is obtained based on the total predicted power generation and load priority. The discrete microgrid is regulated by a global control strategy, and the fault recovery status is continuously monitored. When fault recovery is detected, the main microgrid is smoothly connected to the grid based on its electrical parameters.

[0005] Preferably, the continuous detection of the communication status between the main microgrid and the discrete microgrid via the heartbeat mechanism includes: The communication status is determined by the bidirectional interaction delay and response success status of the periodic heartbeat messages between the main microgrid and the discrete microgrid.

[0006] Preferably, obtaining the global control strategy based on the fault category includes: If the fault type is a low-frequency, low-voltage fault caused by an external fault, the power deficit is obtained, and a global control strategy is obtained in combination with load priority. If the fault type is a low-frequency, low-voltage fault caused by an internal fault, the faulty section is located and isolated for topology reconstruction, and the power deficit after reconstruction is obtained, and a global control strategy is obtained in combination with load priority.

[0007] Preferably, the step of constructing a long-term energy prediction model based on the maintenance cycle through LSTM and residual connections includes: The number of power prediction sub-models is determined based on the maintenance cycle, and a power prediction sub-model is constructed based on LSTM. The gating update frequency, number of hidden layers, and time step are obtained according to the temporal granularity corresponding to the power prediction sub-model. The output of the hidden layer in the power prediction sub-model is added to the input of the hidden layer through residual jump connection. The attention mechanism is embedded in the unified fully connected layer of each power prediction sub-model before the power prediction sub-model is fused to obtain a long-term power prediction model.

[0008] Preferably, the step of training the long-term electricity prediction model using historical photovoltaic power generation sequences and historical meteorological sequences includes: Historical meteorological sequences are aggregated based on temporal granularity to obtain multi-temporal training samples; Multiple time-series training samples are input into the base layer of the long-time-series power prediction model to obtain time-series feature vectors. The time-series feature vectors are then weighted and fused through the long-time-series fusion layer of the long-time-series power prediction model to obtain weighted time-series feature vectors. The weighted time-series feature vector is input into the regression mapping layer in the long-time-series electricity prediction model to obtain the photovoltaic power generation prediction sequence; The power loss sequence and cumulative power loss are obtained by subtracting the photovoltaic power generation sequence from the photovoltaic power generation sequence item by item. The error trend loss is obtained based on the power loss sequence. The waveform change trend loss of the photovoltaic power generation sequence and the photovoltaic power generation sequence is obtained. The total loss value is obtained based on the cumulative power loss, error trend loss, and waveform change trend loss. The parameters in the long-term power prediction model are then corrected based on the total loss value.

[0009] Preferably, the step of obtaining the total predicted power generation of photovoltaic power during the maintenance cycle through a long-term power prediction model, and obtaining a global control strategy based on the total predicted power generation and load priority, includes: Real-time meteorological sequences are obtained based on the correspondence between historical meteorological sequences and historical weather forecast data; The real-time meteorological sequence is aggregated according to the time series granularity and then input into the long-time series power prediction model to obtain the predicted power generation of photovoltaic power in each period during the maintenance cycle. The predicted power generation is then summarized to obtain the total predicted power generation. The sum of the total predicted power generation, the total amount of electricity stored in energy storage devices, and the total amount of electricity to be generated by the diesel generator units is obtained. Using this sum as a hard constraint, a global control strategy is obtained by considering load priority and predicted power generation.

[0010] Preferably, before performing power consumption regulation on the discrete microgrid according to the global control strategy, the method further includes: The system actively establishes the voltage and frequency reference of the discrete microgrid by issuing operation mode switching commands to the energy storage equipment, and simultaneously issues black start commands to the diesel generator unit. After adjusting the output voltage, frequency and phase of the diesel generator unit to match the discrete microgrid bus, the diesel generator unit is connected to the discrete microgrid bus.

[0011] Preferably, the smooth grid connection based on the electrical parameters of the main microgrid includes: If the discrete microgrid is in synchronous recovery mode, the electrical parameters of the main microgrid are obtained, and the electrical parameters of the power supply in the discrete microgrid are adjusted. When the electrical parameters of the main microgrid and the electrical parameters of the power supply meet the synchronous parallel conditions, the grid connection switch is closed. If the discrete microgrid is in power outage recovery mode, the load switch and power switch in the discrete microgrid are disconnected, and the grid connection switch is closed.

[0012] Preferably, the smooth grid connection based on the electrical parameters of the main microgrid further includes: When the difference between the voltage amplitude in the electrical parameters of the main microgrid and the voltage amplitude in the electrical parameters of the power supply is less than a first preset difference, and the difference between the frequency in the electrical parameters of the main microgrid and the frequency in the electrical parameters of the power supply is less than a second preset difference, and the difference between the phase angle in the electrical parameters of the main microgrid and the phase angle in the electrical parameters of the power supply is less than a third preset difference, it indicates that the conditions for simultaneous parallel operation are met.

[0013] By adopting the above technical solution, the present invention has the following advantages: (1) When communication is disconnected, a long-term power prediction model is constructed by using maintenance cycle, LSTM, and residual connection. This overcomes the drawback of insufficient accuracy when using a single LSTM for long-term power prediction. The total predicted power generation of photovoltaics within the maintenance cycle is obtained based on the long-term power prediction model, and a global control strategy is obtained by combining load priority. This solves the technical problem that existing technologies cannot effectively regulate discrete microgrids under long-term power and communication disconnection. This eliminates the dependence on the master station or central controller and ensures that discrete microgrids still have complete self-healing capabilities under conditions of complete communication disconnection. (2) Considering that LSTM is prone to gradient vanishing when processing multi-step time series data, as the prediction time increases, deep networks struggle to transmit early meteorological time series features, leading to a continuous amplification of long-term photovoltaic power generation prediction errors. Therefore, within each power generation prediction sub-model, the hidden layer output and hidden layer input are directly added through residual jump connections, thereby reducing the gradient decay defect of deep time series networks, ensuring the complete transmission of long-term time series features, and improving the accuracy of photovoltaic power generation prediction during the maintenance cycle. Considering the differences in meteorological data density and fluctuation characteristics at different time series granularities, a single LSTM is difficult to adapt to long-term photovoltaic power generation prediction scenarios, and is prone to the problem of feature fitting deviation accumulating continuously with the prediction time and prediction results gradually drifting, resulting in a significant decrease in long-term prediction accuracy. Therefore, a long-term time-series power generation prediction model is constructed. For each sub-model, specific gating update frequency, number of hidden layers, time step, and weight parameters are adaptively matched based on the time-series scale characteristics. The model autonomously learns the nonlinear mapping relationship between time-series features of different granularities and photovoltaic power generation. This effectively avoids mutual interference caused by the superposition of features of different scales and automatically optimizes the weight ratio of each granularity feature according to operating conditions, adaptively tilting towards effective time-series features. The construction of this long-term time-series power generation prediction model fundamentally suppresses the error accumulation and result drift problems of long-term prediction, improving the accuracy and stability of long-term photovoltaic power generation prediction for island microgrids. (3) By monitoring the fault recovery status in real time throughout the process, after the fault is cleared, the electrical parameters such as voltage amplitude, frequency and phase of the main microgrid are collected, and the electrical parameters of the discrete microgrid are adjusted to keep it synchronized with the main microgrid before the grid connection operation is performed. This flexible and smooth synchronous grid connection method can effectively suppress the closing inrush current and avoid impact damage to photovoltaic inverters, energy storage converters and other electrical equipment, thus improving the safety of the entire grid connection and disconnection process.

[0014] This invention also provides a microgrid fault self-healing control system, applicable to the aforementioned microgrid fault self-healing control method, comprising: The communication status detection module continuously detects the communication status between the main microgrid and the discrete microgrid through a heartbeat mechanism. When the communication status is detected as connected, it drives the global control strategy generation module to obtain the global control strategy according to the fault category. When the communication status is detected as disconnected, it drives the global control strategy generation module to build a long-term power prediction model based on the maintenance cycle through LSTM and residual connection. The long-term power prediction model is trained through the photovoltaic historical power generation sequence and historical meteorological sequence. The total predicted power generation of photovoltaics within the maintenance cycle is obtained through the long-term power prediction model. The global control strategy is obtained based on the total predicted power generation and load priority. The grid connection and control module is used to regulate the power consumption of the discrete microgrid according to the global control strategy and continuously monitor the fault recovery status. When fault recovery is detected, it performs smooth grid connection according to the electrical parameters of the main microgrid.

[0015] By adopting the above technical solution, the present invention has the following advantages: When communication is disconnected, a long-term power generation prediction model is constructed by utilizing maintenance cycles, LSTM, and residual connections. This overcomes the shortcomings of insufficient accuracy when using a single LSTM for long-term power generation prediction. The total predicted power generation of photovoltaics within the maintenance cycle is obtained based on the long-term power generation prediction model, and a global control strategy is derived by combining load priority. This solves the technical problem of effectively controlling discrete microgrids under conditions of long-term power and communication outages.

[0016] The present invention also provides a computer device, including: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the microgrid fault self-healing control method.

[0017] The present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, performs the steps of the microgrid fault self-healing control method. Attached Figure Description

[0018] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0019] Figure 1 This is a flowchart illustrating the microgrid fault self-healing control method of the present invention. Figure 2This is a schematic diagram of the training process for the long-term energy prediction model of this invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only one preferred embodiment of this invention and are only used to explain this invention. They do not limit the scope of protection of this invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0021] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations (or steps) can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the figures; the process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0022] Example 1: As Figure 1 As shown, the microgrid fault self-healing control method includes the following steps: S1: Continuously monitor the communication status between the main microgrid and the discrete microgrid through a heartbeat mechanism.

[0023] Understandably, the main microgrid refers to the upper-level control microgrid connected to the municipal distribution network on the mainland side, which is responsible for issuing unified dispatch instructions, while the discrete microgrid is an isolated sub-microgrid deployed independently on the island side. The two rely on submarine fiber optic composite cables to complete power transmission and communication interaction.

[0024] In some embodiments, the continuous detection of the communication status between the main microgrid and the discrete microgrid via the heartbeat mechanism includes: The communication status is determined by the bidirectional interaction delay and response success status of the periodic heartbeat messages between the main microgrid and the discrete microgrid.

[0025] The discrete microgrid sends an uplink heartbeat request message carrying its local timestamp to the main microgrid at fixed intervals. Upon receiving the heartbeat request message, the main microgrid immediately sends a downlink response message carrying its own timestamp. The discrete microgrid records the sending time of the heartbeat request and the receiving time of the response message to calculate the bidirectional interaction delay of a single communication and continuously counts the number of consecutive successful responses. If the bidirectional interaction delay exceeds a preset delay threshold, or if no response message is received from the main microgrid for three consecutive heartbeat cycles, the communication between the discrete microgrid and the main microgrid is considered interrupted; otherwise, the communication between the two is considered to be normal.

[0026] S2: If the communication status is connected, the global control strategy is obtained according to the fault category. If the communication status is disconnected, a long-term power prediction model is constructed based on the maintenance cycle by connecting LSTM and residuals. The long-term power prediction model is trained by the photovoltaic historical power generation sequence and historical meteorological sequence. The total predicted power generation of photovoltaics within the maintenance cycle is obtained through the long-term power prediction model. The global control strategy is obtained based on the total predicted power generation and load priority.

[0027] As is well known, under communication connectivity, discrete microgrids can exchange operational data, fault information, and dispatch commands with the main microgrid in real time. The control logic of the discrete microgrid still needs to comply with the unified management approach of the main microgrid, thus only requiring matching of the global collaborative control strategy corresponding to the fault category. Under communication interruption, the data exchange channels between the two ends are cut off, the discrete microgrid cannot receive dispatch commands from the main microgrid, and switches to an islanded independent operation mode. The control objective shifts to maintaining its own voltage and frequency stability, ensuring continuous and reliable power supply to important loads, and achieving autonomous fault self-healing control of the discrete microgrid. To rationally allocate the total power generated by local photovoltaic power output, energy storage devices, and diesel generators, and to formulate load switching schemes, it is necessary to estimate the total photovoltaic power generation during the entire maintenance period, and then generate a global control strategy adapted to the islanded operating condition. Understandably, the global control strategy is used to define the switching timing of different types of loads in different time periods. When the power supply of the discrete microgrid is insufficient, priority is given to ensuring the continuous power supply to critical loads; when the power supply of the discrete microgrid is sufficient, secondary loads are then considered. Based on the predicted total power generation and combined with the energy storage of the energy storage device and the total power generated by the diesel generator, the load switching timing is dynamically matched, realizing flexible load control, continuously ensuring the stable and reliable operation of the isolated microgrid voltage and frequency, and realizing the autonomous fault self-healing regulation of the discrete microgrid.

[0028] Furthermore, considering the long maintenance cycle, such as constraints from marine weather, tides, the scheduling of specialized repair vessels, and underwater fault point salvage operations, the typical repair time is 2 to 13 days. LSTM is prone to gradient vanishing when processing multi-step long-term time-series data. As the prediction time increases, deep networks struggle to transmit early meteorological time-series characteristics, leading to a continuous amplification of long-term photovoltaic power output prediction errors. Therefore, residual connections are introduced to construct a long-term power prediction model, thereby reducing the gradient decay defect of deep time-series networks, ensuring the complete transmission of long-term time-series characteristics, and improving the accuracy of photovoltaic power generation prediction within the maintenance cycle. Considering the differences in meteorological data density and fluctuation characteristics at different time-series granularities, a single LSTM is insufficient for long-term photovoltaic power prediction scenarios. Therefore, a long-term power prediction model is constructed, and the model autonomously learns the nonlinear mapping relationship between time-series characteristics of different granularities and photovoltaic power generation, further improving the accuracy of photovoltaic power generation prediction within the maintenance cycle. The construction of a long-term power generation prediction model fundamentally suppresses the problems of error accumulation and result drift in long-term prediction, and improves the accuracy and stability of long-term photovoltaic power generation prediction for island microgrids.

[0029] Specifically, obtaining the global control strategy based on the fault category includes: If the fault type is a low-frequency, low-voltage fault caused by an external fault, the power deficit is obtained, and a global control strategy is obtained in combination with load priority. If the fault type is a low-frequency, low-voltage fault caused by an internal fault, the faulty section is located and isolated for topology reconstruction, and the power deficit after reconstruction is obtained, and a global control strategy is obtained in combination with load priority.

[0030] Understandably, external and internal faults are distinguished by the discrete microgrid boundary. External faults occur outside the discrete microgrid, such as main microgrid failures or damaged tie cables, causing only low-frequency and low-voltage conditions within the discrete microgrid. There are no fault points within the discrete microgrid, and topology reconfiguration is unnecessary. Internal faults, on the other hand, occur within the discrete microgrid's internal network structure, power supply, or load branches. These faults directly disrupt the local topology, requiring the faulty section to be located and isolated, network reconfiguration to be completed, and then the post-reconfiguration power deficit to be calculated. Specifically, the three-phase current and zero-sequence current of the branch are collected in real time by the acquisition terminals deployed at each feeder section of the discrete microgrid. When an internal fault occurs in the discrete microgrid, the fault will cause the three-phase current and zero-sequence current to change drastically. The controller extracts the fault transient electrical quantities such as the change amplitude and unbalance pulse from the three-phase current and zero-sequence current waveforms at the moment of the fault. It selects the switch nodes whose fault transient electrical quantities exceed the preset threshold as fault alarm nodes, retrieves the discrete microgrid topology connection information stored locally, and locks the fault section based on the boundary characteristics that there is a fault alarm switch upstream and no fault transient over-limit signal of the adjacent downstream switch. Finally, it issues a trip command to disconnect the sectional switches at both ends of the fault section and electrically isolate the fault area from the normal power supply network.

[0031] Understandably, the active power imbalance in a microgrid is directly manifested as frequency shift, while the reactive power imbalance is manifested as bus voltage drop. Therefore, the active power deficit and reactive power deficit can be solved separately by combining the deviation of measured voltage and frequency from the rated values ​​with the static droop coefficient of the discrete microgrid, and then superimposed to obtain the power deficit of the discrete microgrid.

[0032] Load priority is determined by comprehensively considering multiple indicators, including load importance, load interruptibility, recovery time constraints, and load start-up impact characteristics. Load importance and priority are positively correlated; for example, safety-assured loads such as communication base stations and emergency medical facilities have higher priority than residential water supply and basic lighting loads. The shorter the tolerable power outage duration, the higher the load priority, with priority ranking following: uninterruptible loads > short-term interruptible loads > long-term interruptible loads. The larger the transient impact amplitude of load switching, the higher the load priority. The longer the estimated recovery period, the higher the power supply guarantee priority for the corresponding load. The comprehensive load priority is calculated by weighted fusion of the four evaluation indicators, with the indicator weights decreasing progressively. The weight order is: load importance > load interruptibility > recovery time constraints > load start-up impact characteristics.

[0033] In some embodiments, the construction of a long-term energy prediction model based on maintenance cycles using LSTM and residual connections includes: The number of power prediction sub-models is determined based on the maintenance cycle, and a power prediction sub-model is constructed based on LSTM. The gating update frequency, number of hidden layers, and time step are obtained according to the temporal granularity corresponding to the power prediction sub-model. The output of the hidden layer in the power prediction sub-model is added to the input of the hidden layer through residual jump connection. The attention mechanism is embedded in the unified fully connected layer of each power prediction sub-model before the power prediction sub-model is fused to obtain a long-term power prediction model.

[0034] It should be noted that the gating update frequency, number of hidden layers, time step, and fully connected layers are all internal structures and hyperparameters of the LSTM network. Considering the differences in data density and fluctuation evolution characteristics of meteorological data at different time-series granularities, relying solely on single-scale time-series features is insufficient to simultaneously account for both intraday instantaneous fluctuations and long-term multi-day variations in photovoltaic power generation, making it difficult to guarantee the prediction accuracy of photovoltaic output within the maintenance cycle. In this embodiment, the number of power prediction sub-models is configured differently based on the maintenance cycle. For example, when the maintenance cycle is daily, hourly, semi-daily, and daily power prediction sub-models are preferably configured for collaborative prediction; when the maintenance cycle reaches the weekly level, a weekly power prediction sub-model is added in addition to the hourly, semi-daily, and daily power prediction sub-models. It is understood that the longer the required maintenance cycle, the more power prediction sub-models are needed. Based on this, the specific number of power prediction sub-models can be flexibly adjusted according to the actual prediction accuracy requirements. The gating update frequency, number of hidden layers, and time step are constructed for each power prediction sub-model to process meteorological data at different time-series granularities. Understandably, the smaller the temporal granularity, the richer the high-frequency fluctuations, random noise, and transient changes in the input sequence, resulting in higher data complexity. This necessitates stacking multiple hidden layers to abstract the original meteorological sequence into a higher-order feature representation through progressively nonlinear transformations. Conversely, the larger the temporal granularity, the more macroscopically aggregated the input data, with smoother fluctuations and significant periodic patterns, leading to lower data complexity. Shallow networks can then effectively extract the core features. Continuing to use deep networks on large-granularity data will cause the model capacity to far exceed the data complexity, resulting in a serious risk of overfitting and causing the network to misinterpret residual aggregation errors as valid patterns. Therefore, the larger the temporal granularity corresponding to the electricity prediction sub-model, the fewer the corresponding hidden layers.

[0035] Understandably, the gating update frequency is used to control the refresh cycle of the forget gate, input gate, and output gate of the power prediction sub-model. Therefore, the larger the temporal granularity of the power prediction sub-model, the larger the corresponding gating update frequency. The gating update frequency can specifically be the temporal granularity. Taking a 7-day maintenance cycle as an example, differentiated gating update frequencies can be configured for power prediction sub-models with different temporal granularities: the gating update frequency of the hourly power prediction sub-model can be set to 1 hour, enabling it to respond to power output fluctuations from minute to hourly levels and fully capture short-term power generation characteristics; the gating update frequency of the daily power prediction sub-model can be set to 24 hours, causing it to trigger the gating state refresh only at daily time nodes, filtering out non-steady-state noise in the hourly dimension, and focusing on extracting cross-day periodic features. The time step determines the length of the input sequence for the power generation prediction sub-model. Therefore, the larger the temporal granularity of the power generation prediction sub-model, the smaller the time step. The time step also directly determines the length of the backpropagation path. The essence of differentiated time step design is to configure the most suitable suppression strategy for error sources at different time scales. That is, fine-grained models use large samples to suppress instantaneous noise errors, while coarse-grained models use short paths to suppress far-end drift errors. The synergy of these two approaches ultimately improves the accuracy of the model. By constructing separate power generation prediction sub-models, the decoupled extraction of multi-scale temporal features is achieved, avoiding the problem of mutual interference between high-frequency details and low-frequency trend features. In addition, by compressing the backpropagation paths of the daily and weekly power generation prediction sub-models, the propagation length of long-sequence chain derivatives is significantly shortened, effectively alleviating the gradient vanishing problem during long-time-series training. This allows the model to fully learn and retain daily and weekly periodic patterns, fundamentally avoiding the inherent defect of a sharp drop in accuracy during conventional single LSTM prediction across days. Finally, by fusing multi-granularity features to obtain photovoltaic power generation data, the accuracy and robustness of photovoltaic power generation prediction within the maintenance cycle are significantly improved. Understandably, increasing the temporal granularity means that meteorological data from a longer period is aggregated within a single input time step, and the effective information entropy of the input vector grows superlinearly with the temporal granularity. If the number of hidden units does not increase synchronously with the information density, an information bottleneck will form at the network input, causing a large amount of high-dimensional coupled and correlated information to be truncated and lost during the encoding stage, resulting in structural prediction bias. Therefore, the larger the temporal granularity corresponding to the electricity prediction sub-model, the more hidden units there are in the hidden layer, and the stronger the sub-model's representation and storage capacity for complex nonlinear change patterns such as multi-order fluctuations and coupled correlations within the time series data.

[0036] Considering that a single LSTM is prone to the vanishing gradient problem when processing multi-step time series data, and that deep networks struggle to transmit early meteorological time series characteristics as the prediction duration increases, leading to a continuous amplification of long-term photovoltaic power output prediction errors, this paper proposes adding the output of the previous hidden layer to the input of the current hidden layer in the power prediction sub-model. The specific expression is as follows: , This represents the hidden state output at time t. This represents the hidden state output at time t-1. Let t represent the input vector at time t, which is composed of the meteorological sequence. It is a composite nonlinear transformation function consisting of a forget gate, an input gate, an output gate, and a cell state update. By directly adding the hidden layer output and input through residual jump connections within each power prediction sub-model, the gradient decay defect of deep temporal networks is reduced, thus improving the accuracy of photovoltaic power generation prediction during the maintenance cycle.

[0037] Considering the multi-dimensional features output by each power generation prediction sub-model, their contribution to photovoltaic power generation under different time-series states exhibits significant non-stationarity and heterogeneity, meaning the importance of each feature fluctuates dynamically over time. To accurately characterize this dynamic mapping relationship and achieve effective conversion from features to power generation, a dynamic feature fusion layer based on an attention mechanism is constructed: First, the attention mechanism is used to adaptively calculate the weights of each multi-dimensional feature for the photovoltaic power generation to be predicted, thereby enhancing the significance of important features and suppressing redundant features. Then, the fused features, after weight recalibration, are input into a unified fully connected mapping network, ultimately completing the nonlinear transformation from the high-dimensional feature space to the target power generation regression value.

[0038] like Figure 2 As shown, in some embodiments, training the long-term electricity prediction model using historical photovoltaic power generation sequences and historical meteorological sequences includes: Historical meteorological sequences are aggregated based on temporal granularity to obtain multi-temporal training samples; Multiple time-series training samples are input into the base layer of the long-time-series power prediction model to obtain time-series feature vectors. The time-series feature vectors are then weighted and fused through the long-time-series fusion layer of the long-time-series power prediction model to obtain weighted time-series feature vectors. The weighted time-series feature vector is input into the regression mapping layer in the long-time-series electricity prediction model to obtain the photovoltaic power generation prediction sequence; The power loss sequence and cumulative power loss are obtained by subtracting the photovoltaic power generation sequence from the photovoltaic power generation sequence item by item. The error trend loss is obtained based on the power loss sequence. The waveform change trend loss of the photovoltaic power generation sequence and the photovoltaic power generation sequence is obtained. The total loss value is obtained based on the cumulative power loss, error trend loss, and waveform change trend loss. The parameters in the long-term power prediction model are then corrected based on the total loss value.

[0039] Understandably, the meteorological sequence is specifically composed of meteorological data affecting photovoltaic power generation. For example, if the cloud cover at time t is a and the solar irradiance is b, then the meteorological sequence at time t is (a, b). In the long-term power generation prediction model, the module used to dynamically fuse the output features of sub-models at different time scales is defined as the long-term fusion layer; the module used to map the fused high-dimensional features to the final prediction result is defined as the regression mapping layer; the base layer refers to all components in the long-term power generation prediction model other than the two functional modules mentioned above.

[0040] Taking a weekly maintenance cycle as an example: When training the hourly power prediction sub-model in the long-term power prediction model, the historical meteorological sequence is directly input into the hourly power prediction sub-model to obtain the first time-series feature vector. When training the half-day power prediction sub-model in the long-term power prediction model, the average meteorological sequence of the historical meteorological sequence within 6 hours is obtained, and the average meteorological sequence is input into the half-day power prediction sub-model to obtain the second time-series feature vector. When training the daily power prediction sub-model in the long-term power prediction model, the average meteorological sequence of the historical meteorological sequence within 24 hours is obtained, and the average meteorological sequence is input into the daily power prediction sub-model to obtain the third time-series feature vector. When training the weekly power prediction sub-model in the long-term power prediction model, the average meteorological sequence of the historical meteorological sequence within 168 hours is obtained, and the average meteorological sequence is input into the weekly power prediction sub-model to obtain the fourth time-series feature vector. After weighted fusion of the first, second, third, and fourth time-series feature vectors through the long-term fusion layer, the result is input into the regression mapping layer to obtain the photovoltaic power generation prediction sequence. Understandably, the average meteorological series is composed of the mean values ​​of the same characteristic dimension parameter from multiple meteorological series.

[0041] The cumulative power loss is obtained by summing each term in the power loss sequence. The error trend loss is used to assess whether the error accumulates and amplifies over time; a larger error trend loss indicates that the error accumulates and amplifies over time. Specifically, a linear regression is performed on the power loss sequence to obtain the slope of the line representing the loss over time. If the slope is positive, it indicates that the loss is gradually increasing, and the prediction becomes less accurate as time progresses; a negative slope or close to 0 indicates no significant drift. By performing linear trend fitting on the time-by-time power loss sequence, the slope of the loss over time is extracted as the error trend loss, subjecting the model to an explicit constraint during training that the prediction error must not accumulate and amplify unidirectionally over time. This loss term directly addresses the most typical far-end drift problem in long-term predictions, i.e., the closer to the end of the maintenance cycle, the more easily the error is amplified due to chain propagation. By minimizing the error trend loss, the positive accumulation of error in the temporal direction is effectively curbed, enabling the model to maintain stable and balanced prediction accuracy throughout the maintenance cycle, especially improving the reliability of power predictions in the latter half of the maintenance cycle. Waveform trend loss is used to measure the consistency between the predicted direction of power generation change and the actual direction of power generation change. For example, if the historical photovoltaic power generation sequence shows a waveform of high power generation in the morning, low power generation at noon, and high power generation in the evening, while the predicted photovoltaic power generation sequence shows a waveform of low power generation in the morning, high power generation at noon, and low power generation in the evening, although the point-by-point numerical error may be small, the error in the direction of change will directly lead to errors in the discharge strategy of energy storage devices and the judgment of diesel generator start-up and shutdown timing, thus affecting the power supply reliability of discrete microgrids. Specifically, the first-order difference between the predicted photovoltaic power generation sequence and the historical photovoltaic power generation sequence is obtained to obtain the predicted change sequence and the actual change sequence. The consistency of the signs of the predicted change and the actual change at each moment is compared. When the signs are different, it indicates that the model has made an error in predicting the direction of power generation change. The proportion of the number of moments with different signs is counted out of the total number of moments, and this proportion is used as the waveform trend loss. Taking a photovoltaic (PV) power generation prediction sequence of (50, 80, 100, 110) and a historical PV power generation sequence of (50, 72, 88, 80) as an example, the predicted change sequence is (30, 20, 10), while the actual change sequence is (22, 16, -8). In this case, the waveform trend loss is one-third. This waveform trend loss, by penalizing directional prediction errors, forces the model's predicted sequence to maintain consistency with the actual physical process at the temporal evolution trend level. This effectively avoids scheduling malfunctions caused by similar values ​​but opposite directions, improving the timeliness and accuracy of islanded microgrid control strategies. The parameters in the long-term power generation prediction model are corrected based on the total loss value, enabling the model to accurately learn the mapping relationship between multi-scale fusion features and PV power generation, ultimately improving the accuracy of long-term PV power generation prediction.

[0042] As one embodiment, the step of obtaining the total predicted power generation of photovoltaic power during the maintenance cycle through a long-term time-series power prediction model, and obtaining a global control strategy based on the total predicted power generation and load priority, includes: Real-time meteorological sequences are obtained based on the correspondence between historical meteorological sequences and historical weather forecast data; The real-time meteorological sequence is aggregated according to the time series granularity and then input into the long-time series power prediction model to obtain the predicted power generation of photovoltaic power in each period during the maintenance cycle. The predicted power generation is then summarized to obtain the total predicted power generation. The sum of the total predicted power generation, the total amount of electricity stored in energy storage devices, and the total amount of electricity to be generated by the diesel generator units is obtained. Using this sum as a hard constraint, a global control strategy is obtained by considering load priority and predicted power generation.

[0043] As is well known, the weather forecast data within the maintenance period can be compared with historical weather forecast data using a dynamic time warping algorithm. The historical meteorological sequence corresponding to the closest historical weather forecast data is used as the real-time meteorological data within the maintenance period. Specifically, the real-time meteorological data within the maintenance period refers to hourly meteorological data throughout the entire maintenance period. For example, if the maintenance period is one day, specifically January 1, 2026, then the real-time meteorological data is the 24-hour meteorological data for January 1, 2026. The acquisition of weather forecast data within the maintenance period does not depend on real-time data transmission during communication interruptions. It is adaptively provided by the meteorological data service unit deployed locally in the discrete microgrid based on the actual communication status: when the communication link is connected, the meteorological data service unit receives and caches the weather forecast data sent by the main microgrid in real time; when the communication link is disconnected, it calls the cached weather forecast data in the meteorological data service unit for inference, without relying on external real-time communication.

[0044] In this embodiment, setting the total electricity consumption equal to the total power generation during the maintenance cycle as a hard constraint aims to ensure that the total electricity consumption of the discrete microgrid as a whole does not exceed the total electricity that the discrete microgrid can produce during the maintenance cycle, thus avoiding a continuous power deficit. A global coordinated control strategy is generated through predicted electricity generation for each time period, aiming to allocate flexible loads with electricity consumption adjustment capabilities to periods of sufficient photovoltaic output according to load priority levels, thereby maximizing the utilization of photovoltaic clean energy.

[0045] S3: Perform power consumption regulation on the discrete microgrid according to the global control strategy and continuously monitor the fault recovery status. When fault recovery is detected, perform smooth grid connection according to the electrical parameters of the main microgrid.

[0046] In some embodiments, before performing power consumption regulation on the discrete microgrid according to the global control strategy, the method further includes: The system actively establishes the voltage and frequency reference of the discrete microgrid by issuing operation mode switching commands to the energy storage equipment, and simultaneously issues black start commands to the diesel generator unit. After adjusting the output voltage, frequency and phase of the diesel generator unit to match the discrete microgrid bus, the diesel generator unit is connected to the discrete microgrid bus.

[0047] A V / f operating mode switching command is sent to the converter in the energy storage device. The energy storage device completes the mode switching within 100ms, establishing the islanded grid voltage and frequency reference in a constant voltage and frequency manner. A black start command is simultaneously sent to all available diesel generator units, monitoring status parameters such as speed, oil pressure, and water temperature during the diesel generator startup process. When multiple diesel generators start in parallel, the unit that first reaches the islanded grid voltage and frequency is used as the main power source. After the diesel generator starts, the output voltage, frequency, and phase are synchronized with the islanded grid bus established by the energy storage device by adjusting the diesel generator speed governor and excitation system. When the output voltage, frequency, and phase of the diesel generator meet the synchronization conditions (voltage difference <5%, frequency difference <0.1Hz, phase angle difference <5°) with the voltage, frequency, and phase of the islanded grid bus, the diesel generator output switch is closed to establish the connection between the diesel generator and the islanded grid bus. If the first diesel generator fails to start, it will automatically switch to the backup diesel generator or extend the support time of the energy storage device within 3 seconds. It should be noted that after each level of load is restored, the frequency and voltage on the islanded bus must be checked for stability before restoring the next level of load. Once the communication link is restored, all operational data of the discrete microgrid during the communication interruption period, such as voltage / frequency curves, load change records, power output records, and protection action records, must be uploaded in batches from the secure access area to the distribution substation area 1.

[0048] In some embodiments, the smooth grid connection based on the electrical parameters of the main microgrid includes: If the discrete microgrid is in synchronous recovery mode, the electrical parameters of the main microgrid are obtained, and the electrical parameters of the power supply in the discrete microgrid are adjusted. When the electrical parameters of the main microgrid and the electrical parameters of the power supply meet the synchronous parallel conditions, the grid connection switch is closed. If the discrete microgrid is in power outage recovery mode, the load switch and power switch in the discrete microgrid are disconnected, and the grid connection switch is closed.

[0049] Specifically, the smooth grid connection based on the electrical parameters of the main microgrid also includes: When the difference between the voltage amplitude in the electrical parameters of the main microgrid and the voltage amplitude in the electrical parameters of the power supply is less than a first preset difference, and the difference between the frequency in the electrical parameters of the main microgrid and the frequency in the electrical parameters of the power supply is less than a second preset difference, and the difference between the phase angle in the electrical parameters of the main microgrid and the phase angle in the electrical parameters of the power supply is less than a third preset difference, it indicates that the conditions for simultaneous parallel operation are met.

[0050] Understandably, if the energy storage devices or diesel generators in the discrete microgrid are operating stably and the voltage and frequency of the islanded grid bus are within the normal range, the synchronization recovery mode is preferred to minimize power outages to the loads. If all power sources in the islanded grid have shut down or the voltage and frequency have significantly deviated from the normal range, the power outage recovery mode is selected. In synchronization recovery mode, the voltage amplitude, frequency, and phase angle of the main grid side at the grid connection point are collected. The output voltage amplitude, frequency, and phase angle of the main power source (energy storage or diesel generator) in the discrete microgrid are adjusted to ensure that its deviation from the main grid side meets the synchronization conditions (voltage difference <5%, frequency difference <0.1Hz, phase angle difference <5°). After meeting the synchronization conditions, the grid connection switch is closed, and the discrete microgrid is connected to the main microgrid with zero inrush current. In power outage recovery mode: all load switches and distributed power source switches in the discrete microgrid are disconnected. The grid connection switch between the discrete microgrid and the main microgrid is closed, and the main microgrid supplies power to the discrete microgrid bus. After detecting that the discrete microgrid bus voltage is stable, the load power supply is gradually restored according to the global control strategy.

[0051] When a discrete microgrid is scheduled to disconnect from the grid, i.e. when the main microgrid power source needs maintenance, the dispatcher sets the planned disconnection time window and islanding operation duration through the distribution substation and executes the following steps: obtain the power balance margin during islanding operation; if the margin is insufficient, automatically generate a pre-charging strategy for energy storage devices or a controllable load pre-allocation strategy; through pre-preparation, ensure that the source and load are basically balanced at the time of disconnection; at the planned disconnection time, prioritize adjusting the output of energy storage devices to make the power exchange at the grid connection point approach zero, achieve seamless disconnection under zero power exchange conditions, reduce the impact when the grid connection switch is disconnected; after disconnection, the energy storage devices automatically switch to V / f mode, monitor the islanding frequency and voltage in real time, and dynamically adjust the charging and discharging power of energy storage and the output of diesel generators according to load changes and photovoltaic power fluctuations to maintain the power balance of the islanding grid.

[0052] Example 2: This example also provides a microgrid fault self-healing control system, applicable to the aforementioned microgrid fault self-healing control method, including: The communication status detection module continuously detects the communication status between the main microgrid and the discrete microgrid through a heartbeat mechanism. When the communication status is detected as connected, it drives the global control strategy generation module to obtain the global control strategy according to the fault category. When the communication status is detected as disconnected, it drives the global control strategy generation module to build a long-term power prediction model based on the maintenance cycle through LSTM and residual connection. The long-term power prediction model is trained through the photovoltaic historical power generation sequence and historical meteorological sequence. The total predicted power generation of photovoltaics within the maintenance cycle is obtained through the long-term power prediction model. The global control strategy is obtained based on the total predicted power generation and load priority. The grid connection and control module is used to regulate the power consumption of the discrete microgrid according to the global control strategy and continuously monitor the fault recovery status. When fault recovery is detected, it performs smooth grid connection according to the electrical parameters of the main microgrid.

[0053] Example 3: This example also provides a computer device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the microgrid fault self-healing control method.

[0054] Example 4: This example also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the microgrid fault self-healing control method.

[0055] The specific embodiments described above are preferred embodiments of the microgrid fault self-healing control method, system and equipment of the present invention, and are not intended to limit the specific scope of the present invention. The scope of the present invention includes but is not limited to the specific embodiments described above. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A microgrid fault self-healing control method, characterized in that, Includes the following steps: The communication status between the main microgrid and the discrete microgrids is continuously monitored through a heartbeat mechanism; If the communication status is connected, a global control strategy is obtained based on the fault category. If the communication status is disconnected, a long-term power prediction model is constructed based on the maintenance cycle through LSTM and residual connections. The long-term power prediction model is trained through historical photovoltaic power generation sequences and historical meteorological sequences. The total predicted power generation of photovoltaics within the maintenance cycle is obtained through the long-term power prediction model. A global control strategy is obtained based on the total predicted power generation and load priority. The discrete microgrid is regulated by a global control strategy, and the fault recovery status is continuously monitored. When fault recovery is detected, the main microgrid is smoothly connected to the grid based on its electrical parameters. The long-term energy prediction model based on maintenance cycles, constructed by connecting LSTM and residuals, includes: The number of power prediction sub-models is determined based on the maintenance cycle, and power prediction sub-models are constructed based on LSTM. The gating update frequency, number of hidden layers and time step are obtained according to the time granularity corresponding to the power prediction sub-model. The output of the hidden layer in the power prediction sub-model is added to the input of the hidden layer through residual jump connection. The attention mechanism is embedded in the unified fully connected layer of each power prediction sub-model before the power prediction sub-models are fused to obtain a long time series power prediction model. The training of the long-term energy prediction model using historical photovoltaic power generation sequences and historical meteorological sequences includes: Historical meteorological sequences are aggregated based on temporal granularity to obtain multi-temporal training samples; Multiple time-series training samples are input into the base layer of the long-time-series power prediction model to obtain time-series feature vectors. The time-series feature vectors are then weighted and fused through the long-time-series fusion layer of the long-time-series power prediction model to obtain weighted time-series feature vectors. The weighted time-series feature vector is input into the regression mapping layer in the long-time-series electricity prediction model to obtain the photovoltaic power generation prediction sequence; The power loss sequence and cumulative power loss are obtained by subtracting the photovoltaic power generation sequence from the photovoltaic power generation sequence item by item. The error trend loss is obtained based on the power loss sequence. The waveform change trend loss of the photovoltaic power generation sequence and the photovoltaic power generation sequence is obtained. The total loss value is obtained based on the cumulative power loss, error trend loss, and waveform change trend loss. The parameters in the long-term power prediction model are then corrected based on the total loss value.

2. The microgrid fault self-healing control method according to claim 1, characterized in that, The continuous monitoring of the communication status between the main microgrid and the discrete microgrids via the heartbeat mechanism includes: The communication status is determined by the bidirectional interaction delay and response success status of the periodic heartbeat messages between the main microgrid and the discrete microgrid.

3. The microgrid fault self-healing control method according to claim 1, characterized in that, The process of obtaining the global control strategy based on the fault category includes: If the fault type is a low-frequency, low-voltage fault caused by an external fault, the power deficit is obtained, and a global control strategy is obtained in combination with load priority. If the fault type is a low-frequency, low-voltage fault caused by an internal fault, the faulty section is located and isolated for topology reconstruction, and the power deficit after reconstruction is obtained, and a global control strategy is obtained in combination with load priority.

4. The microgrid fault self-healing control method according to claim 1, characterized in that, The process involves obtaining the total predicted power generation of photovoltaic power during the maintenance cycle through a long-term time-series power prediction model, and then obtaining a global control strategy based on the total predicted power generation and load priority, including: Real-time meteorological sequences are obtained based on the correspondence between historical meteorological sequences and historical weather forecast data; The real-time meteorological sequence is aggregated according to the time series granularity and then input into the long-time series power prediction model to obtain the predicted power generation of photovoltaic power in each period during the maintenance cycle. The predicted power generation is then summarized to obtain the total predicted power generation. The sum of the total predicted power generation, the total amount of electricity stored in energy storage devices, and the total amount of electricity to be generated by the diesel generator units is obtained. Using this sum as a hard constraint, a global control strategy is obtained by considering load priority and predicted power generation.

5. The microgrid fault self-healing control method according to claim 1, characterized in that, Before performing power consumption regulation on the discrete microgrid according to the global control strategy, the following steps are also included: The system actively establishes the voltage and frequency reference of the discrete microgrid by issuing operation mode switching commands to the energy storage equipment, and simultaneously issues black start commands to the diesel generator unit. After adjusting the output voltage, frequency and phase of the diesel generator unit to match the discrete microgrid bus, the diesel generator unit is connected to the discrete microgrid bus.

6. The microgrid fault self-healing control method according to claim 1, characterized in that, The smooth grid connection based on the electrical parameters of the main microgrid includes: If the discrete microgrid is in synchronous recovery mode, the electrical parameters of the main microgrid are obtained, and the electrical parameters of the power supply in the discrete microgrid are adjusted. When the electrical parameters of the main microgrid and the electrical parameters of the power supply meet the synchronous parallel conditions, the grid connection switch is closed. If the discrete microgrid is in power outage recovery mode, the load switch and power switch in the discrete microgrid are disconnected, and the grid connection switch is closed.

7. The microgrid fault self-healing control method according to claim 6, characterized in that, The smooth grid connection based on the electrical parameters of the main microgrid also includes: When the difference between the voltage amplitude in the electrical parameters of the main microgrid and the voltage amplitude in the electrical parameters of the power supply is less than the first preset difference, and the difference between the frequency in the electrical parameters of the main microgrid and the frequency in the electrical parameters of the power supply is less than the second preset difference, and the difference between the phase angle in the electrical parameters of the main microgrid and the phase angle in the electrical parameters of the power supply is less than the third preset difference, it indicates that the conditions for simultaneous parallel operation are met.

8. A microgrid fault self-healing control system, applicable to the microgrid fault self-healing control method according to any one of claims 1-7, characterized in that, include: The communication status detection module continuously detects the communication status between the main microgrid and the discrete microgrid through a heartbeat mechanism. When the communication status is detected as connected, it drives the global control strategy generation module to obtain the global control strategy according to the fault category. When the communication status is detected as disconnected, it drives the global control strategy generation module to build a long-term power prediction model based on the maintenance cycle through LSTM and residual connection. The long-term power prediction model is trained through the photovoltaic historical power generation sequence and historical meteorological sequence. The total predicted power generation of photovoltaics within the maintenance cycle is obtained through the long-term power prediction model. The global control strategy is obtained based on the total predicted power generation and load priority. The grid connection and control module is used to regulate the power consumption of the discrete microgrid according to the global control strategy and continuously monitor the fault recovery status. When fault recovery is detected, it performs smooth grid connection according to the electrical parameters of the main microgrid.

9. A computer device, characterized in that, include: The computer device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the microgrid fault self-healing control method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, performs the steps of the microgrid fault self-healing control method as described in any one of claims 1-7.

Images