Emergency power supply system multi-source collaborative scheduling optimization method based on deep learning

Abstract

The application provides a kind of based on deep learning's emergency power supply system multi-source collaborative scheduling optimization method, it is related to power system scheduling technical field, including obtaining the real-time operating state of multiple power supply sources and load demand data;Through calculating supply-demand power constraint and operation constraint, an initial scheduling sequence is obtained;Real-time operating state is decomposed according to time scale by wavelet transform, energy-time feature map is constructed, comprehensive complementary coefficient is calculated by complementary discriminant space mapping, and optimization weight matrix is generated;According to the optimization weight matrix, the initial scheduling sequence is corrected to form a collaborative scheduling instruction;Based on fault sample set, construct fault evolution path atlas, calculate risk assessment value and generate priority order;Construct execution time gradient sequence, generate switching control sequence and execute.The application improves the response speed of emergency power supply and the accuracy of multi-source collaborative scheduling, reduces the risk of system failure.

Description

Technical Field

[0001] This invention relates to the field of power system dispatching technology, and in particular to a multi-source collaborative dispatching optimization method for emergency power supply systems based on deep learning. Background Technology

[0002] With the rapid development of the social economy and the continuous growth of electricity demand, emergency power supply systems are playing an increasingly important role in ensuring power supply for critical infrastructure, important locations, and emergencies. Emergency power supply systems typically consist of multiple power sources, such as diesel generators, energy storage batteries, photovoltaic power generation, and wind power generation. How to achieve efficient coordinated scheduling of these multi-source power supply devices to ensure rapid response and stable power supply in emergencies is a key technical problem that urgently needs to be solved in the current power system field.

[0003] Traditional emergency power supply dispatching methods primarily rely on pre-set rules and experience models, switching power sources and allocating power manually or semi-automatically. While these methods can meet basic needs in simple scenarios, their limitations have become increasingly apparent as emergency power supply systems expand in scale and power source types diversify. In recent years, with the development of artificial intelligence and deep learning technologies, researchers have begun to explore applying intelligent algorithms to the dispatching optimization of emergency power supply systems to improve the system's intelligence level and response efficiency.

[0004] However, existing technologies still have some shortcomings and deficiencies. Existing emergency power supply dispatching methods focus on power balance issues at a single time scale, lacking in-depth analysis of the operating characteristics of different power sources at multiple time scales, and making it difficult to fully explore the complementary characteristics between power sources. When making dispatching decisions, they lack the ability to predict and assess potential system failure risks, lack dynamic tracking and analysis mechanisms for the failure evolution process, and cannot identify key risk nodes that may lead to system failure in advance, affecting power supply reliability and system security. Existing dispatching command execution mechanisms lack adaptive adjustment capabilities, failing to dynamically adjust the execution order and response time of commands according to the actual operating status and risk level, resulting in delays in the execution of key commands in high-risk scenarios. Summary of the Invention

[0005] This invention provides a deep learning-based multi-source collaborative scheduling optimization method for emergency power supply systems, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a deep learning-based multi-source collaborative scheduling optimization method for emergency power supply systems, comprising:

[0007] Acquire real-time operating status and load demand data from multiple power sources;

[0008] Based on real-time operating status and load demand data, an initial scheduling sequence is obtained by calculating supply and demand power constraints and operating constraints;

[0009] The real-time operating status is decomposed by wavelet transform according to the time scale. Based on the decomposition results, an energy-time feature map is constructed. The comprehensive complementarity coefficient between power sources is calculated by complementary discrimination spatial mapping to generate an optimized weight matrix.

[0010] Based on the optimized weight matrix, the initial scheduling sequence is corrected to form a collaborative scheduling instruction;

[0011] Based on the set of fault samples in historical operation data, a feature propagation link is constructed, a fault evolution path map is established, a risk assessment value is calculated based on the fault evolution path map, and a priority order of scheduling instructions is generated according to the risk assessment value.

[0012] Based on priority, the coordinated scheduling instructions are divided into multiple instruction gradient groups, an execution time gradient sequence is constructed, and an adaptive response timing framework is formed to generate a power supply switching control sequence.

[0013] Execute the switching control sequence and record the execution results, then perform iterative optimization based on the results.

[0014] In one optional embodiment, the real-time operating status is decomposed by wavelet transform according to the time scale, an energy-time feature map is constructed based on the decomposition results, and the comprehensive complementarity coefficient between power sources is calculated through complementary discriminant spatial mapping to generate an optimized weight matrix, including:

[0015] The real-time running status is divided into multiple time scale levels according to the time window length. Wavelet transform decomposition is performed on the running status data in each time scale level. The state fluctuation component and the state trend component are distinguished by a preset frequency threshold to obtain the state decomposition results of each time scale level.

[0016] Based on the state decomposition results, a two-dimensional energy-time feature map of the power source at each time scale level is constructed. The gradient change rate of adjacent data points in the time axis direction is calculated, and the distribution entropy value of each energy interval in the energy axis direction is calculated. The dynamic change feature vector is generated by combining these features.

[0017] The dynamic change feature vectors of each pair of power sources at the same time scale are mapped to the complementary discrimination space, and the cosine value of the angle between the mapped vectors is calculated to obtain the quantitative index of the degree of complementarity.

[0018] The fusion weight is determined based on the load fluctuation intensity and historical scheduling deviation of the corresponding time period at each time scale level. The quantification index of the complementarity of each pair of power sources at all time scale levels is multiplied and accumulated with the fusion weight to obtain the comprehensive complementarity coefficient.

[0019] The comprehensive complementarity coefficients between all power sources are used to construct an optimized weight matrix.

[0020] In one optional embodiment, based on a set of fault samples from historical operational data, a feature propagation chain is constructed, and a fault evolution path map is established, including:

[0021] Extract a set of fault samples with the time of fault occurrence marked from historical operation data. Backtrack a preset time window with the time of fault occurrence as the end time to obtain the sequence of operation parameters within the corresponding time window. Based on the sequence of operation parameters, identify the upstream and downstream devices directly connected to the faulty device to obtain the set of related devices.

[0022] For the operating parameter sequence of each device in the associated device set, calculate the correlation coefficient with the operating parameter sequence of the faulty device, determine the time delay parameter based on the time difference corresponding to the maximum value of the correlation coefficient, construct feature tuples using the operating parameter sequence, time delay parameter and correlation coefficient, and sort the feature tuples in ascending order of time delay parameter to obtain the feature propagation link;

[0023] By utilizing the feature propagation links of all fault samples, we extract the common combinations of device parameters to determine key nodes. We then connect all feature propagation links through these key nodes to construct a directed graph. With device states as nodes and state transitions as edges, and the frequency of occurrence of feature propagation links in the fault sample set as edge weights, we form a fault evolution path graph.

[0024] In one optional embodiment, the risk assessment value is calculated based on the fault evolution path map, and the priority order for generating scheduling instructions based on the risk assessment value includes:

[0025] Obtain the node set and directed edge weights in the fault evolution path graph, construct the failure propagation matrix, calculate the node importance based on the degree centrality of the nodes and the number of paths between nodes, and perform iterative calculations using the node importance and the failure propagation matrix to obtain the failure propagation matrix.

[0026] Collect real-time operating parameters, calculate the similarity value between the real-time operating parameters and the node features in the fault evolution path map, extract nodes with similarity values ​​exceeding a preset similarity threshold to form a set of primary active nodes, use the failure propagation matrix to perform propagation calculation on the set of primary active nodes to obtain a set of secondary active nodes, and combine the set of primary active nodes and the set of secondary active nodes to form an active node sequence.

[0027] For each node in the activated node sequence, the node failure probability is obtained by multiplying the node importance, the corresponding value of the failure propagation matrix, and the similarity value. The correspondence between node failure and the degree of impact on power supply load is established, the load impact loss caused by each node failure is calculated, and the risk assessment value of the power supply area is calculated based on the node failure probability and the load impact loss.

[0028] The priority order of generating dispatch instructions is based on the risk assessment value of the power supply area.

[0029] In one alternative embodiment, the adaptive response timing framework includes:

[0030] Construct a priority distribution curve based on the priority values ​​of each scheduling instruction in the priority order. Based on the priority distribution curve, extract the inflection point position with the largest curvature change to determine the time gradient division boundary, and divide the scheduling instructions into multiple instruction gradient groups.

[0031] For each instruction gradient group, calculate the state stability and switching preparation time of the power supply within the group, and sum the switching preparation time with the state stability as the weight to obtain the group execution delay. Construct an execution time gradient sequence based on the execution delay of each group.

[0032] Establish a coupling relationship matrix between adjacent gradients in the execution time gradient sequence, record the mutual influence coefficient of power supply switching between each execution time gradient, calculate the minimum safe interval between adjacent execution time gradients based on the coupling relationship matrix, and verify and adjust the execution time gradient sequence.

[0033] When the deviation of the load demand data from the historical average exceeds the preset deviation boundary value, the deviation is mapped to a framework adjustment factor. The group execution latency is updated and the execution time series is reconstructed based on the framework adjustment factor to form an adaptive response timing framework.

[0034] In one optional embodiment, when the deviation of the load demand data from the historical average exceeds a preset deviation boundary value, the deviation is mapped to a frame adjustment factor. The group execution latency is updated and the execution time series is reconstructed based on the frame adjustment factor to form an adaptive response timing framework, including:

[0035] Obtain the statistical distribution characteristics of historical load demand data, calculate the deviation of load demand data from the historical mean, and when the deviation exceeds the preset deviation boundary value, map the deviation into a frame adjustment factor according to the preset interval.

[0036] The execution delay of the adjustment group is obtained by multiplying the frame adjustment factor with the state stability of the power supply in each instruction gradient group, and the execution time gradient sequence is reconstructed based on the execution delay of the adjustment group.

[0037] Based on the reconstructed execution time gradient sequence, the coupling relationship matrix between adjacent execution time gradients is extracted, and the minimum safe interval between adjacent execution time gradients is calculated according to the mutual influence coefficient of power supply switching in the coupling relationship matrix.

[0038] By using the minimum safety interval as a constraint, the execution time gradient sequence is checked and adjusted to form an adaptive response timing framework that takes into account dynamic load changes.

[0039] In one alternative embodiment, the generation of the power supply switching control sequence includes:

[0040] Within each gradient interval of the execution time gradient sequence, a dynamic response characteristic curve of the power supply is established. Based on the dynamic response characteristic curve, the ramp rate and settling time of the power supply are calculated to obtain the dynamic switching constraints of each power supply.

[0041] For each combination of scheduling instructions within a time gradient, a power supply group switching topology is constructed, the transmission delay and power fluctuation amplitude of each switching path in the topology are calculated, and the optimal switching path with the minimum transmission delay and limited power fluctuation is selected.

[0042] Based on dynamic switching constraints and the optimal switching path, a switching control sequence is generated that includes switching timing, switching objects, and power allocation.

[0043] A second aspect of this invention provides a deep learning-based multi-source collaborative scheduling optimization system for emergency power supply systems, comprising:

[0044] The first unit is used to acquire real-time operating status and load demand data of multiple power supplies;

[0045] The second unit is used to obtain the initial scheduling sequence based on real-time operating status and load demand data by calculating supply and demand power constraints and operating constraints;

[0046] The third unit is used to decompose the real-time operating status by wavelet transform according to the time scale, construct an energy-time feature map based on the decomposition results, calculate the comprehensive complementarity coefficient between power sources through complementary discrimination spatial mapping, and generate an optimized weight matrix.

[0047] The fourth unit is used to correct the initial scheduling sequence based on the optimized weight matrix and form cooperative scheduling instructions;

[0048] The fifth unit is used to construct feature propagation links and establish fault evolution path maps based on a set of fault samples in historical operation data, calculate risk assessment values ​​based on the fault evolution path maps, and generate priority order of scheduling instructions based on the risk assessment values.

[0049] The sixth unit is used to divide the cooperative scheduling instructions into multiple instruction gradient groups according to the priority order, construct the execution time gradient sequence and form an adaptive response timing framework, and generate the power supply switching control sequence.

[0050] The seventh unit is used to execute the switching control sequence and record the execution results, and to perform iterative optimization based on the execution results.

[0051] A third aspect of the present invention provides an electronic device, comprising:

[0052] processor;

[0053] Memory used to store processor-executable instructions;

[0054] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0055] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0056] In this embodiment of the invention, by acquiring real-time operating status and load demand data of multiple power sources, and based on this data calculating power supply and demand constraints and operating constraints to obtain an initial scheduling sequence, it is possible to respond in real time to the dynamic changes in the emergency power supply system's demand, improving the timeliness and accuracy of scheduling decisions. This avoids the problem that traditional static scheduling methods cannot adapt to real-time operating condition changes, ensuring the reliability and stability of emergency power supply. By decomposing the real-time operating status according to the time scale using wavelet transform, an energy time feature map is constructed. Furthermore, by calculating the comprehensive complementarity coefficients between power sources through complementary discriminant spatial mapping, an optimized weight matrix is ​​generated to correct the initial scheduling sequence, fully exploring the potential differences between different power sources. The complementary characteristics of these technologies enable multi-source collaborative optimization configuration, improving energy utilization efficiency, reducing power supply costs, and enhancing the system's adaptability and robustness under complex operating conditions. By constructing a fault evolution path graph to calculate risk assessment values ​​and generating priority order of scheduling instructions based on these values, the collaborative scheduling instructions are divided into multiple instruction gradient groups. An execution time gradient sequence is constructed to form an adaptive response timing framework, realizing intelligent scheduling decisions based on risk prediction. This effectively reduces system operation risks, improves emergency response speed, and continuously improves the scheduling strategy through iterative optimization of execution results, enabling the system to have self-learning and self-optimization capabilities and improving the overall intelligence level of the emergency power supply system. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the multi-source collaborative scheduling optimization method for emergency power supply systems based on deep learning, according to an embodiment of the present invention.

[0058] Figure 2 A flowchart for constructing a fault evolution map based on correlation analysis is provided. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0061] Figure 1 This is a flowchart illustrating the multi-source collaborative scheduling optimization method for emergency power supply systems based on deep learning, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0062] Acquire real-time operating status and load demand data from multiple power sources;

[0063] Based on real-time operating status and load demand data, an initial scheduling sequence is obtained by calculating supply and demand power constraints and operating constraints;

[0064] The real-time operating status is decomposed by wavelet transform according to the time scale. Based on the decomposition results, an energy-time feature map is constructed. The comprehensive complementarity coefficient between power sources is calculated by complementary discrimination spatial mapping to generate an optimized weight matrix.

[0065] Based on the optimized weight matrix, the initial scheduling sequence is corrected to form a collaborative scheduling instruction;

[0066] Based on the set of fault samples in historical operation data, a feature propagation link is constructed, a fault evolution path map is established, a risk assessment value is calculated based on the fault evolution path map, and a priority order of scheduling instructions is generated according to the risk assessment value.

[0067] Based on priority, the coordinated scheduling instructions are divided into multiple instruction gradient groups, an execution time gradient sequence is constructed, and an adaptive response timing framework is formed to generate a power supply switching control sequence.

[0068] Execute the switching control sequence and record the execution results, then perform iterative optimization based on the results.

[0069] In one optional embodiment, real-time operating status data of multiple power sources are acquired through monitoring equipment at each power supply node. This data includes key indicators such as output power, remaining capacity, and operating time. Simultaneously, power demand data from the load side is collected, including load size and fluctuation characteristics. For example, for three power sources in an emergency scenario, data is collected for: Source A: output power 85kW, remaining power 70%, operating for 3.5 hours; Source B: output power 120kW, remaining power 45%, operating for 5.2 hours; Source C: output power 60kW, remaining power 90%, just put into use for 0.5 hours. The load demand is a total power of 215kW, with a fluctuation range of ±15kW.

[0070] Based on the acquired real-time data, power supply and demand constraints and operational constraints are calculated. Power supply and demand constraints ensure that the total output power of the power sources meets the load demand, while operational constraints include maximum output power limits for each source and minimum remaining capacity requirements. A scheduling model is constructed based on these constraints, and an initial scheduling sequence is generated using a priority ranking algorithm. Specifically, considering that source A has moderate remaining power, source B has low remaining power, and source C is almost fully charged, the initial scheduling scheme is as follows: source A maintains an output of 85kW, source B's output is reduced to 70kW, and source C's output is increased to 60kW.

[0071] To further optimize the scheduling scheme, the real-time operating status data of each power source is decomposed using wavelet transform according to the time scale to separate different frequency components. Typically, the db4 wavelet basis is chosen, and the operating data within 8 hours is decomposed into three levels at 5-minute intervals to obtain low-frequency trend components and multiple high-frequency detail components. Based on the decomposition results, an energy-time feature map is constructed, where the horizontal axis represents time, the vertical axis represents energy density, and the color intensity indicates energy distribution characteristics. After decomposing the data in the above case, it was found that source B exhibits larger high-frequency fluctuations, indicating poor stability; while source C has smaller fluctuation amplitudes and is suitable for stable power supply.

[0072] By employing a complementary discriminant space mapping technique, the comprehensive complementarity coefficients between different power sources are calculated. This technique projects the operating characteristics of each source onto a feature space, calculates the complementarity between eigenvectors, and quantitatively evaluates the complementary relationship between power sources. Example of a comprehensive complementarity coefficient matrix: Complementarity coefficient between source A and source B is 0.72, between source A and source C is 0.45, and between source B and source C is 0.83. Based on these complementarity coefficients, an optimized weight matrix is ​​generated to guide the optimization of the scheduling scheme.

[0073] Based on the optimized weight matrix, the initial scheduling sequence is corrected to form a coordinated scheduling instruction. During the correction process, source pairs with high complementarity coefficients are prioritized for adjustment to achieve complementary energy utilization. The corrected scheduling scheme: source A remains at 85kW, source B is reduced to 65kW, and source C is increased to 65kW, achieving better energy complementarity.

[0074] Simultaneously, based on the set of fault samples in the historical operation database, a feature propagation chain is constructed. This process uses deep learning methods to extract the characteristic patterns of system operation before the fault occurs, identify the precursor features of the fault, and establish causal relationships between features. For example, analyzing equipment overheating fault samples reveals that increased output power fluctuations, unstable output voltage, and rising temperature are typical fault propagation chains. Based on these chains, a fault evolution path graph is established, where nodes represent system states and edges represent state transition probabilities. For the current operating state, a risk assessment value is calculated using the graph. In the above case, the risk assessment value of source B is 0.75, higher than that of source A (0.45) and source C (0.25), indicating that source B has a higher risk of failure.

[0075] The priority order of scheduling instructions is generated based on risk assessment values: high-risk scheduling instructions have higher priority and need to be executed faster. Power adjustment instructions for source B have the highest priority, followed by source C, and lastly source A. Coordinated scheduling instructions are divided into multiple instruction gradient groups according to priority; for example, source B adjustment instructions are assigned to the first gradient, source C to the second gradient, and source A to the third gradient.

[0076] An execution time gradient sequence is constructed to form an adaptive response timing framework. The first gradient instruction is executed immediately, the second gradient instruction is executed with a 30-second delay, and the third gradient instruction is executed with a 60-second delay. This timing arrangement allows the system to observe the effect of each adjustment before executing the next step, avoiding system oscillation. Based on the gradient sequence, a detailed switching control sequence is generated, including parameters such as the power adjustment step size and adjustment rate for each source.

[0077] The switching control sequence is executed and the results are recorded, including power change curves and system stability indicators. The deviations between the execution results and the expected targets are analyzed, such as the error between the actual output and the target value, and the overshoot during the adjustment process. Specifically, during the process of reducing the power of source B from 120kW to 65kW, there is a 5% overshoot, reaching a minimum of 61.75kW before recovering to the target value. Using this execution result data, reinforcement learning methods are employed for iterative optimization, continuously improving the effectiveness and robustness of the scheduling strategy.

[0078] In this embodiment, the coordinated and optimized scheduling of multiple power sources in the emergency power supply system is realized. While meeting the load demand, multiple objectives such as complementary energy utilization and fault risk avoidance are considered, which significantly improves the reliability and energy efficiency of the system.

[0079] In one optional implementation, the real-time operating status is decomposed by wavelet transform according to the time scale, an energy-time feature map is constructed based on the decomposition results, and the comprehensive complementarity coefficient between power sources is calculated through complementary discriminant spatial mapping to generate an optimized weight matrix, including:

[0080] The real-time running status is divided into multiple time scale levels according to the time window length. Wavelet transform decomposition is performed on the running status data in each time scale level. The state fluctuation component and the state trend component are distinguished by a preset frequency threshold to obtain the state decomposition results of each time scale level.

[0081] Based on the state decomposition results, a two-dimensional energy-time feature map of the power source at each time scale level is constructed. The gradient change rate of adjacent data points in the time axis direction is calculated, and the distribution entropy value of each energy interval in the energy axis direction is calculated. The dynamic change feature vector is generated by combining these features.

[0082] The dynamic change feature vectors of each pair of power sources at the same time scale are mapped to the complementary discrimination space, and the cosine value of the angle between the mapped vectors is calculated to obtain the quantitative index of the degree of complementarity.

[0083] The fusion weight is determined based on the load fluctuation intensity and historical scheduling deviation of the corresponding time period at each time scale level. The quantification index of the complementarity of each pair of power sources at all time scale levels is multiplied and accumulated with the fusion weight to obtain the comprehensive complementarity coefficient.

[0084] The comprehensive complementarity coefficients between all power sources are used to construct an optimized weight matrix.

[0085] In one specific implementation, real-time operating status data is divided and processed according to the length of a time window. A base time window of 1 minute is set as the smallest time scale, subsequently expanding to five time scales: 5 minutes, 15 minutes, 30 minutes, and 1 hour. For a specific application scenario of a power distribution system in an industrial park, real-time power output data from photovoltaic power generation units, wind power generation units, energy storage units, and diesel power generation units are collected, with a sampling frequency of 10 times per second. 600 sampling points are obtained within a 1-minute time window, and 3000 sampling points are obtained within a 5-minute time window. The collected raw power data is preprocessed, removing outliers exceeding 200% of the normal range, and missing data points are filled by averaging the values ​​of adjacent data.

[0086] Wavelet transform decomposition is performed on the operational status data within each time scale level. The Daubechies wavelet basis function is selected as the decomposition basis, and the number of decomposition levels is adaptively adjusted according to the time scale level: 3 levels for a 1-minute window, 4 levels for a 5-minute window, and 5 levels for a 15-minute window. For the power data sequence of the photovoltaic power generation unit within a 1-minute window, the first level of decomposition decomposes the original signal into low-frequency approximate components and high-frequency detail components. The low-frequency components reflect the stable trend of power output, while the high-frequency components reflect rapid fluctuation characteristics. A frequency threshold of one-tenth of the sampling frequency (i.e., one change per second) is set as the distinguishing criterion. When the dominant frequency of the decomposed component is higher than this threshold, it is classified as a state fluctuation component; when the dominant frequency is lower than this threshold, it is classified as a state trend component. After decomposing the data of the wind power generation unit into 5 layers within a 15-minute window, the main frequencies of the high-frequency detail components in the first three layers are 0.8 times per second, 0.4 times per second, and 0.2 times per second, respectively, all of which are higher than the threshold and are identified as fluctuation components; the main frequencies of the low-frequency components in the last two layers are 0.05 times per second and 0.02 times per second, respectively, and are identified as trend components.

[0087] A two-dimensional energy-time feature map is constructed based on the state trend component and state fluctuation component obtained from the decomposition. The horizontal axis represents the time dimension, and the vertical axis represents the energy dimension. The energy values ​​are power-normalized values, limited to the range of 0 to 1. A one-minute time window is evenly divided into 60 time units, each unit corresponding to a duration of 1 second. The energy axis is divided into 20 energy intervals, each interval having a width of 0.05 units. For the data of the energy storage unit in a certain one-minute window, the energy values ​​of the state trend component over the 60 time units are extracted, which are 0.32, 0.34, 0.35, 0.36, 0.38, 0.39, and continuously changing to 0.65. These energy values ​​are plotted on a two-dimensional plane to form a continuous curve, constituting the energy-time feature map of the power supply at this time scale.

[0088] Calculate the gradient rate of change of adjacent data points along the time axis on the feature map. Select the energy values ​​of two adjacent time units, calculate the difference between the energy value of the next time step and the energy value of the previous time step, and then divide by the time interval to obtain the gradient value. The gradient of the energy storage unit from the 1st second to the 2nd second is 0.34 minus 0.32, divided by 1 second, resulting in 0.02 units per second. Calculate the gradients of all adjacent points sequentially, obtaining a sequence of 59 gradient values. Analyze the positive and negative distribution of the gradient values. A positive gradient of 63% indicates a significant upward trend in energy, while a negative gradient of 37% indicates a phase of energy decline. The average absolute value of the gradient is calculated to be 0.015 units per second, and the standard deviation is 0.008 units per second, reflecting the drastic nature of the gradient change.

[0089] Calculate the distribution entropy value for each energy interval along the energy axis. Statistically analyze the frequency distribution of energy values ​​falling into 20 energy intervals over 60 time units. In the energy storage unit data, there are 12 data points in the energy interval 0.30 to 0.35, 15 data points in the interval 0.35 to 0.40, and 8 data points in the interval 0.60 to 0.65. Calculate the probability of each energy interval: the probability of the first interval is 12 divided by 60, which equals 0.2, and the probability of the second interval is 0.25. Take the natural logarithm of all non-zero probability values, multiply by this probability value, sum all the products, and take the negative value to obtain the distribution entropy value of 2.74. A larger entropy value indicates a more uniform energy distribution and more complex fluctuation characteristics; a smaller entropy value indicates that energy is concentrated in a few intervals, and the operating state is relatively stable.

[0090] A dynamically changing feature vector is generated by combining gradient and entropy features. The five feature parameters—gradient mean, gradient standard deviation, positive gradient percentage, negative gradient percentage, and distribution entropy—are arranged in order to form a five-dimensional feature vector. The feature vector for the energy storage unit on a 1-minute timescale is [0.015, 0.008, 0.63, 0.37, 2.74]. The feature vector for the photovoltaic power generation unit on the same timescale is [0.022, 0.013, 0.58, 0.42, 3.12]. The feature vector for the diesel power generation unit is [0.003, 0.001, 0.51, 0.49, 1.86]. The differences in feature vectors between different power sources reflect the essential differences in their dynamic characteristics.

[0091] The dynamic feature vectors of each power source pair are mapped to a complementary discrimination space. A five-dimensional complementary discrimination space is constructed, with the five axes of the spatial coordinate system corresponding to the five components of the feature vectors. The feature vectors of the energy storage unit and the photovoltaic power generation unit are taken as two vector points in the space. The dot product of the two vectors is calculated by multiplying the corresponding dimension components and summing them. The first dimension of the energy storage unit (0.015) is multiplied by the first dimension of the photovoltaic unit (0.022), and the second dimension (0.008) is multiplied by 0.013. The sum of these five terms yields a dot product value of 0.0126. The magnitudes of the two vectors are calculated separately. The magnitude of the energy storage unit is the square root of the sum of the squares of the components in each dimension, which is 0.0173. The magnitude of the photovoltaic unit is 0.0226. The dot product value is divided by the product of the two magnitudes to obtain a cosine value of 0.322. This value represents a quantitative indicator of the complementarity between the two power sources at this time scale. The closer the cosine value is to negative 1, the stronger the complementarity; the closer it is to positive 1, the stronger the synchronicity.

[0092] The complementarity between energy storage units and diesel generator units was quantitatively calculated. The dot product of their eigenvectors is 0.0089, the modulus of the energy storage unit is 0.0173, the modulus of the diesel generator unit is 0.0112, and the cosine value is -0.458, indicating a strong complementary relationship. The cosine value between the photovoltaic unit and the diesel generator unit is -0.612, showing even more significant complementarity. The cosine value between the wind power unit and the photovoltaic unit is -0.385, indicating a moderate degree of complementarity.

[0093] Determine the fusion weights for each time scale level. Collect load fluctuation data from the past 24 hours and calculate the load fluctuation intensity for each time scale level corresponding to the time period. The load fluctuation intensity at the 1-minute scale is defined as the difference between the maximum and minimum load power during that period divided by the average value, which is calculated to be 0.18. The fluctuation intensity at the 5-minute scale is 0.32, at the 15-minute scale it is 0.45, at the 30-minute scale it is 0.52, and at the 1-hour scale it is 0.61. The larger the fluctuation intensity, the more important the impact of that time scale on scheduling decisions. Extract scheduling deviation data for each time scale from historical scheduling records. The average deviation rate is 8% at the 1-minute scale, 12% at the 5-minute scale, 15% at the 15-minute scale, 18% at the 30-minute scale, and 22% at the 1-hour scale. Multiply the load fluctuation intensity by the reciprocal of the historical scheduling deviation and normalize to obtain the fusion weights for each time scale as 0.12, 0.18, 0.22, 0.24, and 0.24, respectively.

[0094] The complementarity of energy storage and photovoltaic units across all time scales was quantitatively quantified and weighted. The complementarity index for both was 0.322 at the 1-minute scale, 0.287 at the 5-minute scale, -0.156 at the 15-minute scale, -0.289 at the 30-minute scale, and -0.412 at the 1-hour scale. Multiplying each scale's complementarity index by its corresponding fusion weight (0.322 multiplied by 0.12, plus 0.287 multiplied by 0.18), and accumulating these five terms, a comprehensive complementarity coefficient of -0.084 was obtained. The comprehensive complementarity coefficient between energy storage and diesel units was -0.336, between photovoltaic and diesel units was -0.428, and between wind power and photovoltaic units was -0.265.

[0095] The comprehensive complementarity coefficients of all power sources are arranged in pairs to form an optimized weight matrix. Each row and column of the matrix contains four power sources, and diagonal elements are set to zero to indicate no complementarity relationship. The first row and second column contain the comprehensive complementarity coefficient of energy storage and photovoltaic (-0.084); the first row and third column contain the coefficient of energy storage and wind power (-0.192); and the first row and fourth column contain the coefficient of energy storage and diesel power (-0.336). The second row and third column contain the coefficient of photovoltaic and wind power (-0.265); the second row and fourth column contain the coefficient of photovoltaic and diesel power (-0.428); and the third row and fourth column contain the coefficient of wind power and diesel power (-0.381).

[0096] In one optional implementation, based on a set of fault samples from historical operational data, a feature propagation chain is constructed, and a fault evolution path map is established, including:

[0097] Extract a set of fault samples with the time of fault occurrence marked from historical operation data. Backtrack a preset time window with the time of fault occurrence as the end time to obtain the sequence of operation parameters within the corresponding time window. Based on the sequence of operation parameters, identify the upstream and downstream devices directly connected to the faulty device to obtain the set of related devices.

[0098] For the operating parameter sequence of each device in the associated device set, calculate the correlation coefficient with the operating parameter sequence of the faulty device, determine the time delay parameter based on the time difference corresponding to the maximum value of the correlation coefficient, construct feature tuples using the operating parameter sequence, time delay parameter and correlation coefficient, and sort the feature tuples in ascending order of time delay parameter to obtain the feature propagation link;

[0099] By utilizing the feature propagation links of all fault samples, we extract the common combinations of device parameters to determine key nodes. We then connect all feature propagation links through these key nodes to construct a directed graph. With device states as nodes and state transitions as edges, and the frequency of occurrence of feature propagation links in the fault sample set as edge weights, we form a fault evolution path graph.

[0100] In one specific implementation, fault risk assessment is a crucial step in the deep learning-based multi-source collaborative scheduling optimization method for emergency power supply systems. Accurate assessment requires constructing a fault evolution path graph, extracting a set of fault samples with labeled fault occurrence times from a historical operational database. These samples contain the specific time of the fault occurrence and information about the involved equipment. For example, in a certain emergency power supply scenario, 30 sets of fault samples were extracted from the database, including 10 sets of inverter overheating faults, 8 sets of battery management system faults, 7 sets of distribution cabinet connection faults, and 5 sets of control system communication faults. Taking the inverter overheating fault as an example, the fault occurrence time was recorded as 14:32:15 on June 15, 2025.

[0101] Starting from the moment the fault occurs, a preset time window is traced backward to obtain the sequence of operating parameters within that time window. Typically, the tracing time window is set to 30 minutes before the fault occurs, with a sampling interval of 10 seconds, resulting in data acquisition at 180 time points. For inverter overheating faults, the operating parameters of the faulty device (inverter) during these 30 minutes are extracted, including key parameters such as output current, output voltage, internal temperature, cooling fan speed, and ambient temperature. Specifically, the inverter output current gradually increases from 55A to 78A, the output voltage fluctuates from 220V to 210V, the internal temperature rises from 65℃ to 95℃, the cooling fan speed increases from 2000rpm to 3500rpm, and the ambient temperature remains around 38℃.

[0102] Based on the acquired operating parameter sequence, upstream and downstream devices directly connected to the faulty device are identified to determine the associated device set. For inverter overheating faults, upstream devices include DC power supplies (such as battery banks and diesel generators), and downstream devices include distribution cabinets and load controllers. According to the power supply system topology, the associated device set for the inverter is determined to be: battery bank, diesel generator, main distribution cabinet, and load controller. For these associated devices, the operating parameter sequence for the 30 minutes prior to the fault is also extracted. For example, the discharge current, terminal voltage, and temperature of the battery bank; the output current, fuel consumption rate, and speed of the diesel generator; the bus voltage and branch currents of the main distribution cabinet; and the input voltage and load power of the load controller.

[0103] For each device in the associated device set, the correlation coefficient between its operating parameter sequence and the operating parameter sequence of the faulty device is calculated. A sliding time window is used to calculate the correlation, with a window size of 5 minutes and a step size of 30 seconds. By calculating the correlation of parameter sequences under different time windows, the maximum correlation coefficient and its corresponding time difference are identified to determine the time delay parameter. In the inverter overheating fault case, the correlation coefficient between the battery pack discharge current and the inverter internal temperature is found to be 0.92, with a time difference of 4 minutes and 20 seconds corresponding to the maximum correlation point. This indicates that an abnormal battery pack discharge current occurs approximately 4 minutes and 20 seconds earlier than the inverter temperature. The correlation coefficient between the diesel generator output current fluctuation and the inverter output voltage fluctuation is 0.85, with a time difference of 3 minutes and 10 seconds. The correlation coefficient between the main distribution cabinet bus voltage fluctuation and the inverter output current is 0.78, with a time difference of 2 minutes and 50 seconds.

[0104] Feature tuples are constructed using the acquired operating parameter sequence, time delay parameters, and correlation coefficients. The structure of the feature tuples is: <Source device, source parameter, target device, target parameter, time delay parameter, correlation coefficient>. Based on the aforementioned calculation results, the following feature tuples are constructed: <Battery pack, discharge current, inverter, internal temperature, 260 seconds, 0.92>, <Diesel generator, output current, inverter, output voltage, 190 seconds, 0.85>, <Main distribution cabinet, bus voltage, inverter, output current, 170 seconds, 0.78>. These feature tuples are sorted in ascending order of time delay parameters to obtain the feature propagation path: <Abnormal main distribution cabinet bus voltage> → <Diesel generator output current fluctuation> → <Excessive battery pack discharge current> → <Inverter overheating>.

[0105] The above analysis process was repeated for all fault samples to obtain the characteristic propagation links of different types of faults. By analyzing the characteristic propagation links of multiple fault samples, the common combinations of equipment parameters were extracted to identify key nodes. In the emergency power supply system, by analyzing 30 sets of fault samples, the main key nodes identified included: battery pack status (normal / over-discharge / overcharge), inverter status (normal / overheating / overload), distribution cabinet status (normal / overcurrent / short circuit), and controller status (normal / communication interruption / control abnormality).

[0106] All feature propagation paths are connected through key nodes to construct a directed graph. In this directed graph, device states are treated as nodes, state transitions as edges, and the frequency of occurrence of feature propagation paths in the fault sample set is used as edge weights to form a fault evolution path graph. For example, in the path graph obtained from the analysis of 30 sets of fault samples, the path "battery over-discharge → inverter input voltage insufficient → load power supply instability" appeared 6 times in the samples, and its edge weight was set to 6; the path "main distribution cabinet overcurrent → inverter overload → inverter overheating" appeared 8 times, and its edge weight was 8. The entire fault evolution path graph contains multiple possible fault evolution paths, and each path is marked with its corresponding frequency of occurrence.

[0107] Based on the real-time operating status of the current power supply system, it is mapped to nodes in the fault evolution path graph. A depth-first search algorithm is used to find all possible fault evolution paths, and a risk assessment value is calculated based on the edge weights. The risk assessment value calculation considers the cumulative sum of the edge weights along the path, and is normalized by combining the path length. For example, if the power supply system detects a battery bank discharge current of 85A, higher than the normal value of 65A, mapped to the "Battery Bank Over-Discharge" node in the path graph, two possible evolution paths are found: Path 1 "Battery Bank Over-Discharge → Insufficient Inverter Input Voltage → Unstable Load Power Supply" with a risk value of 0.75; Path 2 "Battery Bank Over-Discharge → Battery Management System Overcurrent Protection → System Emergency Shutdown" with a risk value of 0.62. Based on the calculation results, the most likely fault evolution direction of the current system is determined to be unstable load power supply, with a risk assessment value of 0.75.

[0108] Based on the risk assessment results, a priority order is generated for scheduling instructions. Instructions related to high-risk paths receive higher priority and are executed earlier to prevent potential failures. In the above case, for the high-risk path of "battery over-discharge → insufficient inverter input voltage → unstable load power supply," the generated scheduling instructions include reducing the inverter load and starting the backup power supply, and are given the highest priority to ensure these instructions are executed first, thereby effectively preventing failures.

[0109] like Figure 2As shown, a flowchart illustrating the construction process of a fault evolution map based on correlation analysis is presented.

[0110] In one optional implementation, the risk assessment value is calculated based on the fault evolution path graph, and the priority order for generating scheduling instructions based on the risk assessment value includes:

[0111] Obtain the node set and directed edge weights in the fault evolution path graph, construct the failure propagation matrix, calculate the node importance based on the degree centrality of the nodes and the number of paths between nodes, and perform iterative calculations using the node importance and the failure propagation matrix to obtain the failure propagation matrix.

[0112] Collect real-time operating parameters, calculate the similarity value between the real-time operating parameters and the node features in the fault evolution path map, extract nodes with similarity values ​​exceeding a preset similarity threshold to form a set of primary active nodes, use the failure propagation matrix to perform propagation calculation on the set of primary active nodes to obtain a set of secondary active nodes, and combine the set of primary active nodes and the set of secondary active nodes to form an active node sequence.

[0113] For each node in the activated node sequence, the node failure probability is obtained by multiplying the node importance, the corresponding value of the failure propagation matrix, and the similarity value. The correspondence between node failure and the degree of impact on power supply load is established, the load impact loss caused by each node failure is calculated, and the risk assessment value of the power supply area is calculated based on the node failure probability and the load impact loss.

[0114] The priority order of generating dispatch instructions is based on the risk assessment value of the power supply area.

[0115] In one specific implementation, all node information is extracted from the fault evolution path graph. Assume a power supply area contains 50 key nodes, encompassing power facilities such as substations, distribution transformers, switchgear, and line segmentation points. Each node has multi-dimensional characteristic parameters, including equipment voltage level, load capacity, operating years, and historical fault count. Nodes are connected by directed edges, with the edge weight representing the strength of fault propagation. The weight value is set from 0.1 to 1.0. For example, the edge weight from a substation to a downstream distribution transformer is 0.85, indicating that a fault at this substation has an 85% probability of causing anomalies in downstream equipment.

[0116] When constructing the failure propagation matrix, all nodes are arranged in topological order to form a 50×50 two-dimensional array structure. The element in the i-th row and j-th column of the matrix represents the strength of the direct failure impact of node i on node j. If there is no direct connection between node i and node j, the corresponding matrix element is assigned a value of 0; if there is a direct connection, the weight of the corresponding directed edge is filled into that position. For example, if there is a directed edge with a weight of 0.72 between node 3 and node 7, then the element in the 3-th row and 7-th column of the matrix has a value of 0.72. For each node itself, the diagonal elements of the matrix are uniformly set to 1.0.

[0117] The calculation of node importance is based on the characteristics of the network topology. For each node, its in-degree and out-degree are counted. In-degree represents how many upstream nodes point to the node, and out-degree represents how many downstream nodes the node points to. Taking node 15 as an example, this node has 3 in-edges and 7 out-edges, so its total degree is 10. The number of shortest paths that this node participates in is also calculated by traversing all node pairs in the graph and counting the number of shortest paths passing through node 15. Let's assume the result is 120. The degree is normalized by dividing the node's degree by the maximum degree value in the network. Assuming the maximum degree is 15, the degree centrality of node 15 is 10 divided by 15, which gives 0.667. The number of paths is also normalized by dividing the number of paths that the node participates in by the total number of paths in the network. Assuming the total number of paths is 2450, the path contribution is 120 divided by 2450, which gives 0.049. By weighting degree centrality and path contribution with a ratio of 7:3, the initial importance of node 15 is calculated as 0.667 multiplied by 0.7 plus 0.049 multiplied by 0.3, resulting in a value of 0.481.

[0118] The failure propagation matrix is ​​corrected using node importance. Each element in the matrix is ​​multiplied by the importance of the corresponding node. For example, if node 3 has an importance of 0.52, all non-zero elements in the third row are multiplied by 0.52 for updating. The corrected matrix reflects the actual contribution of nodes of different importance to failure propagation. To capture indirect propagation effects, the corrected matrix is ​​iteratively calculated by performing consecutive multiplications with itself. Each iteration represents one level of failure propagation in the network. A convergence condition is set for the iteration process; it stops when the difference between two consecutive iterations is less than 0.001. A stable state is typically reached after 5 to 8 iterations, ultimately yielding a failure propagation matrix that comprehensively considers both direct and indirect propagation.

[0119] The real-time data acquisition module obtains the current operating parameters of each node from the monitoring system, with a sampling frequency set to once every 5 minutes. The acquired parameters include physical quantities such as node voltage, current, power factor, temperature, and vibration amplitude. Taking data at a certain moment as an example, the real-time voltage of node 8 is 10.2 kV, the current is 145 A, and the temperature is 68 degrees Celsius. Comparing these real-time parameters with the historical characteristics of this node in the spectrum, the historical normal operating characteristics of node 8 are: voltage 10.5 kV, current 130 A, and temperature 55 degrees Celsius. The deviation between the real-time values ​​and the characteristic values ​​is calculated: the voltage deviation rate is 10.2 minus 10.5 divided by 10.5, resulting in -0.029; the current deviation rate is 145 minus 130 divided by 130, resulting in 0.115; and the temperature deviation rate is 68 minus 55 divided by 55, resulting in 0.236. After taking the absolute value of the deviation of each parameter, normalization was performed, and the comprehensive similarity was calculated using the Euclidean distance metric. The similarity of node 8 was calculated by subtracting the square root of the sum of squares of each deviation from 1, resulting in a similarity value of 0.82.

[0120] A similarity threshold of 0.75 was set. The similarity calculations for all 50 nodes were iterated, and nodes with similarities below 0.75 were identified as anomalous nodes. Assuming that nodes 8, 12, 23, 35, and 41 were identified with similarities of 0.82, 0.68, 0.71, 0.65, and 0.73 respectively, these nodes constitute the first-level activated node set. Using the numbers of the first-level activated nodes as indices, corresponding row vectors were extracted from the failure propagation matrix. Each row vector represents the degree of failure impact of that activated node on all other nodes. The five extracted row vectors were summed column-by-column to obtain a cumulative influence vector of length 50, where each element represents the total influence intensity on the corresponding node. The secondary activation threshold is set to 0.3. Nodes with cumulative influence vector values ​​exceeding 0.3 are selected. For example, nodes 6, 19, 28, and 33 have cumulative influence values ​​of 0.42, 0.35, 0.51, and 0.38, respectively. These four nodes form the secondary activation node set. The five primary activation nodes are then merged with the four secondary activation nodes to form an activation node sequence containing nine nodes.

[0121] For each node in the activated node sequence, the failure probability is calculated. Taking node 12 as an example, its importance is 0.53, the maximum propagation value at its corresponding position in the failure propagation matrix is ​​0.68, and its real-time similarity is 0.68. Multiplying these three values ​​by 0.53, then by 1 and subtracting 0.68, yields a failure probability of 0.115 for node 12. The failure probabilities are calculated for all nine nodes in the activated sequence, yielding probabilities of 0.115, 0.089, 0.142, 0.073, 0.096, 0.058, 0.127, 0.081, and 0.104, respectively.

[0122] A mapping relationship between node failure and load loss was established, and the load range supplied by each node was determined by analyzing the power grid topology. Node 12 directly or indirectly supplies power to 8 industrial users, 15 commercial users, and 120 residential users. The unit outage loss for each user type was set at RMB 500 per hour for industrial users, RMB 200 per hour for commercial users, and RMB 50 per hour for residential users. The total load loss caused by the failure of node 12 was calculated as 8 x 500 + 15 x 200 + 120 x 50, resulting in a load loss value of RMB 10,000 per hour. The same method was used to calculate the load loss values ​​of the other 8 nodes in the activation sequence, which were RMB 8,500, RMB 12,000, RMB 6,500, RMB 9,200, RMB 7,800, RMB 14,500, RMB 8,900, and RMB 11,200, respectively.

[0123] The risk assessment value of a node is obtained by multiplying the failure probability of each node by its corresponding load loss value. The risk assessment value of node 12 is 0.115 multiplied by 10000, which equals 1150. After calculating the risk values ​​of all 9 nodes, they are sorted in descending order of their numerical values. Assuming the sorted results are as follows: node 23 risk value 2059, node 28 risk value 1841, node 33 risk value 1434, node 12 risk value 1150, node 41 risk value 1011, node 35 risk value 795, node 19 risk value 696, node 8 risk value 534, and node 6 risk value 465.

[0124] The dispatch instruction generation module determines the priority of handling based on the risk assessment values. A first-priority dispatch instruction is generated for node 23, which has the highest risk value. This instruction includes immediately dispatching inspection personnel to the site, activating backup power, and adjusting the load distribution plan. A second-priority instruction is generated for node 28, which has the second highest risk value, requiring equipment status verification to be completed within 30 minutes. The remaining nodes are then assigned priority levels three through nine, forming a complete sequence of dispatch instructions. This ensures that power grid operation and maintenance resources are rationally allocated according to risk levels, achieving precise management of fault prevention and emergency response.

[0125] In one optional implementation, the adaptive response timing framework includes:

[0126] Construct a priority distribution curve based on the priority values ​​of each scheduling instruction in the priority order. Based on the priority distribution curve, extract the inflection point position with the largest curvature change to determine the time gradient division boundary, and divide the scheduling instructions into multiple instruction gradient groups.

[0127] For each instruction gradient group, calculate the state stability and switching preparation time of the power supply within the group, and sum the switching preparation time with the state stability as the weight to obtain the group execution delay. Construct an execution time gradient sequence based on the execution delay of each group.

[0128] Establish a coupling relationship matrix between adjacent gradients in the execution time gradient sequence, record the mutual influence coefficient of power supply switching between each execution time gradient, calculate the minimum safe interval between adjacent execution time gradients based on the coupling relationship matrix, and verify and adjust the execution time gradient sequence.

[0129] When the deviation of the load demand data from the historical average exceeds the preset deviation boundary value, the deviation is mapped to a framework adjustment factor. The group execution latency is updated and the execution time series is reconstructed based on the framework adjustment factor to form an adaptive response timing framework.

[0130] In one specific implementation, all currently pending power supply scheduling instructions are obtained. Each scheduling instruction carries a priority value, which is set to an integer value from 0 to 100. These scheduling instructions are arranged in descending order of priority value to form a priority sequence. Taking a power distribution system in an industrial park as an example, at a certain moment there are 18 scheduling instructions with priority values ​​of 95, 93, 89, 85, 82, 78, 72, 68, 64, 58, 52, 45, 38, 32, 28, 22, 18, and 12.

[0131] A priority distribution curve is constructed with the priority value as the ordinate and the instruction sequence number as the abscissa. This curve reflects the trend of priority value changes with instruction sequence number. For each point on the curve, the system calculates the curvature value formed by the point and its two adjacent points. The curvature value is calculated by taking the target point and one point before and after it, and using the distance between the three points to calculate the reciprocal of the radius of the arc. In the above 18 instruction examples, the curvature values ​​of instruction number 3, instruction number 7, and instruction number 11 are 0.42, 0.38, and 0.35, respectively, which are significantly higher than the curvature values ​​of other positions in the range of 0.05 to 0.12. The system identifies points where the curvature value exceeds a preset threshold of 0.25 as inflection points, thereby determining the time gradient division boundary. Based on the three identified inflection points, the 18 scheduling instructions are divided into four instruction gradient groups, containing 3, 4, 4, and 7 instructions respectively.

[0132] For the first instruction gradient group, which contains three scheduling instructions with priorities of 95, 93, and 89, involving three power supply units, real-time operating status parameters of these three power supply units were collected, including output voltage stability, output current fluctuation amplitude, and internal temperature change rate. For the first power supply, its voltage stability index was calculated by dividing the output voltage deviation by the rated voltage. The measured deviation was 2 volts, and the rated voltage was 220 volts, resulting in a voltage stability score of 0.009. The current fluctuation amplitude was calculated by dividing the difference between the maximum and minimum current over the last 10 seconds by the average current. The measured maximum current was 52 amps, the minimum current was 48 amps, and the average current was 50 amps, resulting in a fluctuation amplitude of 0.08. The temperature change rate was calculated by dividing the temperature change per minute by the current temperature. The measured temperature change was 1.2 degrees Celsius, and the current temperature was 60 degrees Celsius, resulting in a change rate of 0.02. The average of the reciprocals of these three index values ​​after normalization was taken as the state stability, and the state stability of this power supply was calculated to be 0.89. The state stability of the second and third power supplies, calculated using the same method, are 0.92 and 0.86, respectively.

[0133] Further, the switching preparation time for each power supply is obtained. The switching preparation time includes the sum of three parts: the communication delay for the power supply to receive the switching command, the internal relay action time, and the circuit stabilization time. For the first power supply, the measured communication delay is 15 milliseconds, the relay action time is 80 milliseconds, and the circuit stabilization time is 35 milliseconds, for a total switching preparation time of 130 milliseconds. The switching preparation times for the second and third power supplies are 145 milliseconds and 120 milliseconds, respectively. The state stability of each power supply is used as a weight value, multiplied by the corresponding switching preparation time, and summed to obtain the group execution delay. The group execution delay for the first group is calculated as: state stability 0.89 multiplied by 130 milliseconds + 0.92 multiplied by 145 milliseconds + 0.86 multiplied by 120 milliseconds, resulting in 371.7 milliseconds. The group execution delays for the second, third, and fourth groups are calculated using the same method: 428.3 milliseconds, 385.6 milliseconds, and 512.8 milliseconds, respectively. These four group execution delays are arranged sequentially to form an execution time gradient sequence.

[0134] A coupling relationship matrix between adjacent gradients in the execution time gradient sequence is established. This matrix records the degree of mutual influence between power supply switching between adjacent groups. The degree of influence is determined by analyzing the physical connection relationship and electrical parameter matching degree of the power supplies within the two groups. There are two shared bus connections between the first and second groups. The output voltage phase difference between the last power supply of the first group and the first power supply of the second group is detected to be 8 degrees, and the current capacity ratio is 1.3 times. Based on the degree of deviation of the phase difference and capacity ratio from the standard values, the mutual influence coefficient is determined to be 0.15. The mutual influence coefficient between the second and third groups is 0.22, and between the third and fourth groups it is 0.18. The larger the mutual influence coefficient, the stronger the interference of the previous group's execution on the subsequent group, requiring a longer safety interval.

[0135] The minimum safe interval is calculated based on the mutual influence coefficient. The baseline safe interval is set to 50 milliseconds. The actual safe interval is obtained by multiplying the mutual influence coefficient by the baseline value and adding 10% of the execution delay of the previous group. The minimum safe interval between the first and second groups is calculated as the influence coefficient (0.15) multiplied by 50 milliseconds plus 10% of the first group's delay (371.7 milliseconds), resulting in 44.67 milliseconds. The interval between the second and third groups is 53.83 milliseconds, and the interval between the third and fourth groups is 47.56 milliseconds. The actual intervals in the original execution time gradient sequence are checked to ensure they meet the minimum safe interval requirement. If any discrepancies exist, the execution start time of the corresponding group is shifted forward.

[0136] When a significant change in current load demand data is detected, the framework's adaptive adjustment mechanism is triggered. Load demand data includes instantaneous power demand and duration forecasts. The system maintains historical load demand data for the past 30 days, with a calculated historical average of 850 kW. Currently, a sudden increase in load demand to 1150 kW is detected. The deviation is calculated as the difference between the current value and the historical average divided by the historical average, resulting in a deviation of 0.353. A preset deviation boundary value is set at 0.15. If the current deviation exceeds this boundary value, the system initiates the adjustment process.

[0137] The deviation is mapped to a frame adjustment factor. The mapping uses a piecewise linear function. When the deviation is between 0.15 and 0.30, the adjustment factor is the deviation minus 0.15 multiplied by 2. When the deviation exceeds 0.30, the adjustment factor is 0.30 plus 0.5 times the excess. The current deviation of 0.353 is mapped to an adjustment factor of 0.30 plus 0.053 multiplied by 0.5, resulting in 0.3265. This adjustment factor is used to update the execution latency of each group, multiplying the original group's execution latency by the adjustment factor plus 1. The updated latency for the first group is 371.7 milliseconds multiplied by 1.3265, resulting in 493.0 milliseconds. The remaining groups are updated in the same way to 568.2 milliseconds, 511.5 milliseconds, and 680.3 milliseconds, respectively. Based on the updated group execution latency, the mutual influence coefficient and minimum safety interval are recalculated to construct a new execution time series, completing the dynamic adjustment of the adaptive response timing frame.

[0138] In one optional implementation, when the deviation of the load demand data from the historical average exceeds a preset deviation boundary value, the deviation is mapped to a frame adjustment factor. The group execution latency is updated and the execution time series is reconstructed based on the frame adjustment factor to form an adaptive response timing framework, including:

[0139] Obtain the statistical distribution characteristics of historical load demand data, calculate the deviation of load demand data from the historical mean, and when the deviation exceeds the preset deviation boundary value, map the deviation into a frame adjustment factor according to the preset interval.

[0140] The execution delay of the adjustment group is obtained by multiplying the frame adjustment factor with the state stability of the power supply in each instruction gradient group, and the execution time gradient sequence is reconstructed based on the execution delay of the adjustment group.

[0141] Based on the reconstructed execution time gradient sequence, the coupling relationship matrix between adjacent execution time gradients is extracted, and the minimum safe interval between adjacent execution time gradients is calculated according to the mutual influence coefficient of power supply switching in the coupling relationship matrix.

[0142] By using the minimum safety interval as a constraint, the execution time gradient sequence is checked and adjusted to form an adaptive response timing framework that takes into account dynamic load changes.

[0143] In one specific implementation, the construction process of the adaptive response timing framework begins with the statistical analysis of historical load demand data. Hourly load demand data from the past thirty days is collected, including multi-dimensional parameters such as active power demand, reactive power demand, and voltage quality requirements. After normalizing this historical data, statistical characteristics such as the mean, standard deviation, skewness, and kurtosis of each dimension parameter are calculated. For example, the historical average active power demand is 1500 kW, and the standard deviation is 200 kW. The currently monitored load demand is 2100 kWh. The deviation of this value from the historical mean is obtained by dividing the difference between the two by the standard deviation, specifically 600 kW divided by 200 kW, resulting in a deviation value of 3.

[0144] A pre-set deviation boundary value of 2 is used. When the calculated deviation of 3 exceeds this boundary value, the frame adjustment mechanism is activated. The deviation mapping process employs a piecewise linear mapping strategy, dividing the deviation value range into multiple intervals. When the deviation is between 0 and 1, the frame adjustment factor is set to 1.0, indicating no adjustment is needed; when the deviation is between 1 and 2, the frame adjustment factor linearly increases to 1.2; when the deviation is between 2 and 3, the frame adjustment factor increases from 1.2 to 1.5; and when the deviation exceeds 3, the frame adjustment factor is set to 1.8. According to this mapping rule, the frame adjustment factor corresponding to the current deviation of 3 is 1.5.

[0145] The state stability assessment of power supplies within each instruction gradient group needs to comprehensively consider factors such as the historical switching success rate, response time fluctuation range, and current operating temperature. The first gradient group includes the main transformer, with a historical switching success rate of 98%, a response time standard deviation of 0.3 seconds, a current operating temperature of 60 degrees Celsius, and a rated temperature of 85 degrees Celsius. Through weighted comprehensive calculation, the state stability of this power supply is 0.92. The state stability of the standby transformer in the second gradient group is 0.85, and the state stability of the emergency generator set in the third gradient group is 0.78. Multiplying the framework adjustment factor of 1.5 by the state stability of each gradient group yields an execution delay of 1.38 for the first gradient group, 1.275 for the second gradient group, and 1.17 for the third gradient group.

[0146] The original execution time gradient sequence was set to 0.5 seconds for the first gradient, 1.2 seconds for the second gradient, and 2.0 seconds for the third gradient. After updating with adjusted execution delays, the execution time of the first gradient was adjusted to 0.5 seconds multiplied by 1.38 (0.69 seconds), the execution time of the second gradient was adjusted to 1.2 seconds multiplied by 1.275 (1.53 seconds), and the execution time of the third gradient was adjusted to 2.0 seconds multiplied by 1.17 (2.34 seconds). The reconstructed execution time gradient sequence forms a new time node distribution.

[0147] Extracting the coupling relationship between adjacent execution time gradients requires establishing a coupling relationship matrix. The coupling relationship between the first and second gradients is quantified by analyzing the voltage sag depth, frequency fluctuation amplitude, and phase jump angle during the switching process between the two power supplies. Actual monitoring data shows that when switching from the first gradient main transformer to the second gradient standby transformer, the voltage sag depth is 15%, the frequency fluctuation amplitude is 0.3 Hz, and the phase jump angle is 8 degrees. These parameters constitute the elements in the first row and second column of the coupling relationship matrix, and their mutual influence coefficient is calculated to be 0.25. During the switching process between the second and third gradients, the voltage sag depth is 22%, the frequency fluctuation amplitude is 0.5 Hz, and the phase jump angle is 12 degrees, with a corresponding mutual influence coefficient of 0.38.

[0148] The minimum safety interval is calculated based on the product of the mutual influence coefficient and the time required for the power system to recover to steady state. The reference time for the power system to recover from a disturbance state to steady state is set to 1.0 second. The minimum safety interval between the first and second gradients is the product of the mutual influence coefficient of 0.25, the reference recovery time of 1.0 second, and the safety margin coefficient of 1.5, resulting in 0.375 seconds. The minimum safety interval between the second and third gradients is 0.38 multiplied by 1.0 second and then by 1.5, yielding 0.57 seconds.

[0149] The verification and adjustment process for the execution time gradient sequence checks each reconstructed time node. The execution time of the first gradient is 0.69 seconds, and the execution time of the second gradient is 1.53 seconds, with a time difference of 0.84 seconds. This interval is greater than the minimum safe interval of 0.375 seconds, satisfying the constraint and requiring no adjustment. The time difference between the execution time of the second gradient (1.53 seconds) and the execution time of the third gradient (2.34 seconds) is 0.81 seconds, which is greater than the minimum safe interval of 0.57 seconds, also satisfying the constraint. If the interval between adjacent gradients is less than the minimum safe interval, the execution time of subsequent gradients is shifted forward by the difference between the minimum safe interval and the actual interval.

[0150] The adjusted execution time gradient sequence contains the precise execution time points of all gradients and safety intervals between adjacent gradients. This sequence serves as the core data structure of the adaptive response timing framework, guiding the timing of power supply switching commands in real time. As load demand continuously changes, the deviation is periodically recalculated and the framework adjustment factor is updated to maintain dynamic synchronization between the timing framework and the actual load. This entire adaptive mechanism ensures that, under conditions of drastic load fluctuations, the switching actions of each gradient's power supply can respond quickly to changes in demand while maintaining the stability and continuity of the power system operation.

[0151] In one optional implementation, the power supply switching control sequence includes:

[0152] Within each gradient interval of the execution time gradient sequence, a dynamic response characteristic curve of the power supply is established. Based on the dynamic response characteristic curve, the ramp rate and settling time of the power supply are calculated to obtain the dynamic switching constraints of each power supply.

[0153] For each combination of scheduling instructions within a time gradient, a power supply group switching topology is constructed, the transmission delay and power fluctuation amplitude of each switching path in the topology are calculated, and the optimal switching path with the minimum transmission delay and limited power fluctuation is selected.

[0154] Based on dynamic switching constraints and the optimal switching path, a switching control sequence is generated that includes switching timing, switching objects, and power allocation.

[0155] In one specific implementation, during the process of establishing the dynamic response characteristic curve of the power supply within each gradient interval of the execution time gradient sequence, real-time data on the output power variation over time is collected for each power supply. For example, with three power supplies: power supply A supplies power to the main grid, power supply B is a diesel generator set, and power supply C is an energy storage system. With a time gradient interval set to 5 seconds, a step response test is performed on power supply A. When the load suddenly increases by 100 kW, the entire process of its output power rising from the initial value to the target value is recorded. Power values ​​are sampled every 0.1 seconds, and 50 data points are continuously collected to form a time-power data sequence. These data points are plotted on a coordinate system, with the horizontal axis representing time and the vertical axis representing output power values. The dynamic response characteristic curve of the power supply is obtained by fitting the data points.

[0156] The ramp rate of the power supply is calculated based on the established dynamic response characteristic curve. Data from the power increase phase of the curve is extracted, and the ratio of the power change to the time interval between two adjacent sampling points is calculated. For power supply A, during the test, the power increases from 0 kW to 45 kW between the first and second seconds, with a ramp rate of 45 kW per second; between the second and third seconds, the power increases from 45 kW to 78 kW, with a ramp rate of 33 kW per second. By iterating through all time intervals of the entire increase process, the minimum ramp rate value is taken as the ramp rate index for the power supply. Therefore, the ramp rate of power supply A is ultimately determined to be 33 kW per second.

[0157] When calculating the settling time, the target power value is identified from the response characteristic curve, and the allowable fluctuation range is set to ±2% of the target value. The settling time is the time difference between the moment the step signal is applied and the moment the power enters the stable range, after the power output enters this fluctuation range and remains within it. Power supply A has a target power of 100 kW and an allowable fluctuation range of 98 kW to 102 kW. The curve data shows that the power reaches 98.5 kW at 3.8 seconds and remains within this range; therefore, the settling time for power supply A is 3.8 seconds. The same test and calculation process is performed on power supplies B and C, resulting in a ramp rate of 25 kW / s and a settling time of 5.2 seconds for power supply B, and a ramp rate of 80 kW / s and a settling time of 1.5 seconds for power supply C.

[0158] When obtaining the dynamic switching constraints for each power source, switching rules are formulated by combining the ramp rate and settling time parameters. The dynamic switching constraints for power source A stipulate that the power change per second during switching must not exceed 33 kW, and a settling time of at least 3.8 seconds must be reserved after the switching is completed before the next switching operation can proceed. The dynamic switching constraints for power source B stipulate that the power change per second must not exceed 25 kW, and a settling time of 5.2 seconds must be reserved. The dynamic switching constraints for power source C stipulate that the power change per second must not exceed 80 kW, and a settling time of 1.5 seconds must be reserved. These constraints serve as hard limitations for subsequent switching path planning.

[0159] When constructing a power supply group switching topology graph for scheduling command combinations within each time gradient, all possible switching paths are represented graphically. Assume three scheduling commands are received within a 5-second time gradient: the first command requires a total output power of 150 kW, the second command requires 220 kW, and the third command requires 180 kW. Each power supply is represented as a node in the graph at different power output states. Power supply A provides power from 0 to 200 kW, and is divided into five state nodes: 0 kW, 50 kW, 100 kW, 150 kW, and 200 kW. Power supply B provides power from 0 to 150 kW, and is divided into four state nodes: 0 kW, 50 kW, 100 kW, and 150 kW. Power supply C provides power from 0 to 100 kW, and is divided into three state nodes: 0 kW, 50 kW, and 100 kW. Connection edges are established between the state nodes based on reachability and power matching relationships, forming a multi-level switching topology graph.

[0160] When calculating the transmission delay of each switching path in the topology diagram, the time required for each switching action on the path is accumulated. Switching from the initial state to a state combination satisfying the first instruction of 150 kW, Path 1 selects power supply A to provide 100 kW and power supply C to provide 50 kW. Based on the settling time of power supply A (3.8 seconds) and power supply C (1.5 seconds), and since the two power supplies start in parallel, the transmission delay of this path is taken as the larger value of 3.8 seconds. Path 2 selects power supply A to provide 150 kW, involving only a single switching of power supply A, with a transmission delay of 3.8 seconds. Continuing to calculate the delay of switching from the 150 kW state to the 220 kW state, Path 3 adds power supply B to Path 1, providing 70 kW. Since the ramp rate of power supply B is 25 kW per second, it takes 2.8 seconds to ramp from 0 kW to 70 kW. Adding the settling time of 5.2 seconds, this switching action takes 8 seconds, and the cumulative transmission delay of Path 3 is 11.8 seconds.

[0161] When calculating the power fluctuation amplitude of each switching path, the maximum difference between the total output power and the target power during the switching process is analyzed. In Path 1, power supply A and power supply C start simultaneously. Power supply A outputs 60 kW in the first 2 seconds, and power supply C outputs 40 kW in the first second. At this time, the total output power is 100 kW, which is 50 kW less than the target of 150 kW. The power fluctuation amplitude of this path is 50 kW. In Path 2, only power supply A ramps up, outputting 66 kW in the first 2 seconds, which is 84 kW less than the target of 150 kW. The power fluctuation amplitude of this path is 84 kW. A power fluctuation limit threshold of 60 kW is set. Path 1 meets the constraint, and Path 2 exceeds the limit and is excluded.

[0162] When selecting the optimal switching path with the minimum transmission delay and limited power fluctuation, all paths that meet the power fluctuation constraint are sorted by transmission delay. Among the paths that meet the first instruction, path one has a transmission delay of 3.8 seconds and a power fluctuation of 50 kW. Path four, which selects a combination of power supply B providing 100 kW and power supply C providing 50 kW, has a transmission delay of 5.2 seconds and a power fluctuation of 45 kW. After comparison, path one is selected as the optimal path for the first stage. Subsequent switching paths are then evaluated to form a complete sequence of optimal switching paths.

[0163] When generating the switching control sequence based on dynamic switching constraints and the optimal switching path, the switching timing, switching targets, and power allocation are organized in a structured data format. The first switching instruction is executed at time 0, switching targets are power supply A and power supply C, and the power allocation is power supply A outputting 100 kW and power supply C outputting 50 kW. The second switching instruction is executed at time 5, switching target is power supply B, and the power allocation is adjusted so that power supply A remains at 100 kW, power supply C remains at 50 kW, and power supply B outputs 70 kW. The third switching instruction is executed at time 15, switching target is power supply B, and the power allocation is adjusted so that power supply A remains at 100 kW, power supply C remains at 50 kW, and power supply B decreases to 30 kW. This switching control sequence is issued as control instructions to each power supply controller at the execution layer.

[0164] The multi-source collaborative scheduling optimization system for emergency power supply based on deep learning, as described in this embodiment of the invention, includes:

[0165] The first unit is used to acquire real-time operating status and load demand data of multiple power supplies;

[0166] The second unit is used to obtain the initial scheduling sequence based on real-time operating status and load demand data by calculating supply and demand power constraints and operating constraints;

[0167] The third unit is used to decompose the real-time operating status by wavelet transform according to the time scale, construct an energy-time feature map based on the decomposition results, calculate the comprehensive complementarity coefficient between power sources through complementary discrimination spatial mapping, and generate an optimized weight matrix.

[0168] The fourth unit is used to correct the initial scheduling sequence based on the optimized weight matrix and form cooperative scheduling instructions;

[0169] The fifth unit is used to construct feature propagation links and establish fault evolution path maps based on a set of fault samples in historical operation data, calculate risk assessment values ​​based on the fault evolution path maps, and generate priority order of scheduling instructions based on the risk assessment values.

[0170] The sixth unit is used to divide the cooperative scheduling instructions into multiple instruction gradient groups according to the priority order, construct the execution time gradient sequence and form an adaptive response timing framework, and generate the power supply switching control sequence.

[0171] The seventh unit is used to execute the switching control sequence and record the execution results, and to perform iterative optimization based on the execution results.

[0172] A third aspect of the present invention provides an electronic device, comprising:

[0173] processor;

[0174] Memory used to store processor-executable instructions;

[0175] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0176] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0177] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0178] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based multi-source collaborative scheduling optimization method for emergency power supply systems, characterized in that, include: Acquire real-time operating status and load demand data from multiple power sources; Based on real-time operating status and load demand data, an initial scheduling sequence is obtained by calculating supply and demand power constraints and operating constraints; The real-time operating status is decomposed by wavelet transform according to the time scale. Based on the decomposition results, an energy-time feature map is constructed. The comprehensive complementarity coefficient between power sources is calculated by complementary discrimination spatial mapping to generate an optimized weight matrix. Based on the optimized weight matrix, the initial scheduling sequence is corrected to form a collaborative scheduling instruction; Based on the set of fault samples in historical operation data, a feature propagation link is constructed, a fault evolution path map is established, a risk assessment value is calculated based on the fault evolution path map, and a priority order of scheduling instructions is generated according to the risk assessment value. Based on priority, the coordinated scheduling instructions are divided into multiple instruction gradient groups, an execution time gradient sequence is constructed, and an adaptive response timing framework is formed to generate a power supply switching control sequence. Execute the switching control sequence and record the execution results, then perform iterative optimization based on the results.

2. The method according to claim 1, characterized in that, The real-time operating status is decomposed using wavelet transform on a time scale. Based on the decomposition results, an energy-time feature map is constructed. The comprehensive complementarity coefficient between power sources is calculated through complementary discriminant spatial mapping, and an optimized weight matrix is ​​generated, including: The real-time running status is divided into multiple time scale levels according to the time window length. Wavelet transform decomposition is performed on the running status data in each time scale level. The state fluctuation component and the state trend component are distinguished by a preset frequency threshold to obtain the state decomposition results of each time scale level. Based on the state decomposition results, a two-dimensional energy-time feature map of the power source at each time scale level is constructed. The gradient change rate of adjacent data points in the time axis direction is calculated, and the distribution entropy value of each energy interval in the energy axis direction is calculated. The dynamic change feature vector is generated by combining these features. The dynamic change feature vectors of each pair of power sources at the same time scale are mapped to the complementary discrimination space, and the cosine value of the angle between the mapped vectors is calculated to obtain the quantitative index of the degree of complementarity. The fusion weight is determined based on the load fluctuation intensity and historical scheduling deviation of the corresponding time period at each time scale level. The quantification index of the complementarity of each pair of power sources at all time scale levels is multiplied and accumulated with the fusion weight to obtain the comprehensive complementarity coefficient. The comprehensive complementarity coefficients between all power sources are used to construct an optimized weight matrix.

3. The method according to claim 1, characterized in that, Based on a set of fault samples from historical operational data, a feature propagation chain is constructed, and a fault evolution path map is established, including: Extract a set of fault samples with the time of fault occurrence marked from historical operation data. Backtrack a preset time window with the time of fault occurrence as the end time to obtain the sequence of operation parameters within the corresponding time window. Based on the sequence of operation parameters, identify the upstream and downstream devices directly connected to the faulty device to obtain the set of related devices. For the operating parameter sequence of each device in the associated device set, calculate the correlation coefficient with the operating parameter sequence of the faulty device, determine the time delay parameter based on the time difference corresponding to the maximum value of the correlation coefficient, construct feature tuples using the operating parameter sequence, time delay parameter and correlation coefficient, and sort the feature tuples in ascending order of time delay parameter to obtain the feature propagation link; By utilizing the feature propagation links of all fault samples, we extract the common combinations of device parameters to determine key nodes. We then connect all feature propagation links through these key nodes to construct a directed graph. With device states as nodes and state transitions as edges, and the frequency of occurrence of feature propagation links in the fault sample set as edge weights, we form a fault evolution path graph.

4. The method according to claim 1, characterized in that, Based on the fault evolution path graph, risk assessment values ​​are calculated. The priority order for generating scheduling instructions based on these risk assessment values ​​includes: Obtain the node set and directed edge weights in the fault evolution path graph, construct the failure propagation matrix, calculate the node importance based on the degree centrality of the nodes and the number of paths between nodes, and perform iterative calculations using the node importance and the failure propagation matrix to obtain the failure propagation matrix. Collect real-time operating parameters, calculate the similarity value between the real-time operating parameters and the node features in the fault evolution path map, extract nodes with similarity values ​​exceeding a preset similarity threshold to form a set of primary active nodes, use the failure propagation matrix to perform propagation calculation on the set of primary active nodes to obtain a set of secondary active nodes, and combine the set of primary active nodes and the set of secondary active nodes to form an active node sequence. For each node in the activated node sequence, the node failure probability is obtained by multiplying the node importance, the corresponding value of the failure propagation matrix, and the similarity value. The correspondence between node failure and the degree of impact on power supply load is established, the load impact loss caused by each node failure is calculated, and the risk assessment value of the power supply area is calculated based on the node failure probability and the load impact loss. The priority order of generating dispatch instructions is based on the risk assessment value of the power supply area.

5. The method according to claim 1, characterized in that, The adaptive response timing framework includes: Construct a priority distribution curve based on the priority values ​​of each scheduling instruction in the priority order. Based on the priority distribution curve, extract the inflection point position with the largest curvature change to determine the time gradient division boundary, and divide the scheduling instructions into multiple instruction gradient groups. For each instruction gradient group, calculate the state stability and switching preparation time of the power supply within the group, and sum the switching preparation time with the state stability as the weight to obtain the group execution delay. Construct an execution time gradient sequence based on the execution delay of each group. Establish a coupling relationship matrix between adjacent gradients in the execution time gradient sequence, record the mutual influence coefficient of power supply switching between each execution time gradient, calculate the minimum safe interval between adjacent execution time gradients based on the coupling relationship matrix, and verify and adjust the execution time gradient sequence. When the deviation of the load demand data from the historical average exceeds the preset deviation boundary value, the deviation is mapped to a framework adjustment factor. The group execution latency is updated and the execution time series is reconstructed based on the framework adjustment factor to form an adaptive response timing framework.

6. The method according to claim 5, characterized in that, When the deviation of load demand data from the historical average exceeds a preset deviation boundary value, the deviation is mapped to a framework adjustment factor. Based on this factor, the group execution latency is updated and the execution time series is reconstructed, forming an adaptive response timing framework, including: Obtain the statistical distribution characteristics of historical load demand data, calculate the deviation of load demand data from the historical mean, and when the deviation exceeds the preset deviation boundary value, map the deviation into a frame adjustment factor according to the preset interval. The execution delay of the adjustment group is obtained by multiplying the frame adjustment factor with the state stability of the power supply in each instruction gradient group, and the execution time gradient sequence is reconstructed based on the execution delay of the adjustment group. Based on the reconstructed execution time gradient sequence, the coupling relationship matrix between adjacent execution time gradients is extracted, and the minimum safe interval between adjacent execution time gradients is calculated according to the mutual influence coefficient of power supply switching in the coupling relationship matrix. By using the minimum safety interval as a constraint, the execution time gradient sequence is checked and adjusted to form an adaptive response timing framework that takes into account dynamic load changes.

7. The method according to claim 5, characterized in that, The power supply switching control sequence includes: Within each gradient interval of the execution time gradient sequence, a dynamic response characteristic curve of the power supply is established. Based on the dynamic response characteristic curve, the ramp rate and settling time of the power supply are calculated to obtain the dynamic switching constraints of each power supply. For each combination of scheduling instructions within a time gradient, a power supply group switching topology is constructed, the transmission delay and power fluctuation amplitude of each switching path in the topology are calculated, and the optimal switching path with the minimum transmission delay and limited power fluctuation is selected. Based on dynamic switching constraints and the optimal switching path, a switching control sequence is generated that includes switching timing, switching objects, and power allocation.

8. A deep learning-based multi-source collaborative scheduling optimization system for emergency power supply systems, used to implement the method of any one of claims 1-7, characterized in that, include: The first unit is used to acquire real-time operating status and load demand data of multiple power supplies; The second unit is used to obtain the initial scheduling sequence based on real-time operating status and load demand data by calculating supply and demand power constraints and operating constraints; The third unit is used to decompose the real-time operating status by wavelet transform according to the time scale, construct an energy-time feature map based on the decomposition results, calculate the comprehensive complementarity coefficient between power sources through complementary discrimination spatial mapping, and generate an optimized weight matrix. The fourth unit is used to correct the initial scheduling sequence based on the optimized weight matrix and form cooperative scheduling instructions; The fifth unit is used to construct feature propagation links and establish fault evolution path maps based on a set of fault samples in historical operation data, calculate risk assessment values ​​based on the fault evolution path maps, and generate priority order of scheduling instructions based on the risk assessment values. The sixth unit is used to divide the cooperative scheduling instructions into multiple instruction gradient groups according to the priority order, construct the execution time gradient sequence and form an adaptive response timing framework, and generate the power supply switching control sequence. The seventh unit is used to execute the switching control sequence and record the execution results, and to perform iterative optimization based on the execution results.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images