Power converter optimal configuration method and system based on scenario clustering for distribution network

Abstract

The application relates to the technical field of power conversion control of power distribution networks, and discloses a power distribution network power converter optimization configuration method and system based on scene clustering. The method comprises the following steps: collecting power converter operation data, calculating a power change rate and an accumulated thermal stress index, and constructing a multi-dimensional feature matrix; obtaining typical conversion scenes through multi-level clustering; evaluating the thermal performance of a standby topology, constructing a scene-topology adaptation matrix; predicting the accumulated amount of thermal stress, quantifying switching benefits and costs, and performing topology switching when the net benefits exceed a threshold value in three consecutive rolling prediction periods. The application improves the thermal stress management accuracy of power conversion equipment under diversified operation scenes and the reliability of topology switching decisions.

Description

Technical Field

[0001] This application relates to the field of power conversion control technology in power distribution networks, and in particular to a method and system for optimizing the configuration of power converters in power distribution networks based on scenario clustering. Background Technology

[0002] In power distribution networks, power converters serve as crucial interfaces between distributed photovoltaic (PV) and energy storage systems and the grid, performing DC-AC or DC-DC power conversion. Their operational reliability directly impacts the power supply quality and economic efficiency of the distribution system. Existing power converters in distribution networks typically employ fixed topology configurations, selecting a single topology scheme such as a two-level full-bridge, a three-level NPC (non-parallel) converter, or an interleaved parallel converter for specific power levels. Under rated operating conditions, this topology achieves high conversion efficiency and acceptable device temperature rise through optimized switching frequency and modulation strategies. To address power scenario changes caused by distribution network load fluctuations, existing technologies monitor input power or output current in real time. When the power exceeds a preset threshold, topology switching is triggered. For example, a two-level topology is used when the power is less than 4kW, switching to a three-level topology when the power is between 4kW and 8kW, and switching to a soft-switching topology when the power is greater than 8kW. This static threshold switching strategy establishes switching rules based on the steady-state loss characteristics of different topologies in each power range.

[0003] The shortcomings of existing technologies lie in the fact that static threshold switching strategies rely solely on instantaneous power to determine topology selection, neglecting the temporal dependence of device thermal stress accumulation and the dynamic characteristics of scenario evolution. This leads to a mismatch between topology configuration and actual thermal stress distribution. On the one hand, distribution network operation scenarios have multi-timescale characteristics: power fluctuates rapidly at the second level, heat accumulates at the minute level, and daily load curves change at the hour level. Threshold switching based solely on instantaneous power cannot distinguish the different threats to device thermal lifespan posed by short-term power pulses and continuous heavy loads. It may maintain a high-loss topology even when the instantaneous power triggers the threshold but historical thermal stress is already close to saturation, or delay switching in scenarios where the instantaneous power has not reached the threshold but the power change rate is drastic, causing a rapid temperature rise. On the other hand, existing threshold switching methods simplify switching decisions to a comparison of power and a fixed threshold, without quantifying the transient thermal stress costs generated by the topology switching process itself. These costs include the off-state voltage stress borne by the device during the turn-off phase, the inrush current stress from the discharge and recharging of the output capacitor during the dead-time phase, and the additional heat loss from inefficient operation during the soft-start phase. These switching costs may accumulate in frequent switching scenarios and offset or even exceed the thermal stress reduction benefits brought by topology optimization, making it difficult to guarantee the economy and reliability of switching decisions. Summary of the Invention

[0004] This application provides a method and system for optimizing the configuration of power converters in power distribution networks based on scenario clustering. It solves the problems in the prior art where the topology configuration of power conversion equipment in power distribution networks cannot adapt to the evolution of scenarios at multiple time scales, the switching decision does not quantify the switching cost, and the lack of a prediction and verification mechanism leads to inaccurate thermal stress control of devices. It improves the accuracy of thermal stress management and the reliability of topology switching decisions of power conversion equipment under diverse operating scenarios.

[0005] Firstly, this application provides a method for optimizing the configuration of power converters in a distribution network based on scenario clustering, the method comprising:

[0006] Step S1: Collect the input voltage, output current, instantaneous conversion power, and junction temperature of each switching device of the power converter, calculate the power change rate and cumulative thermal stress index, and construct a multi-dimensional operating characteristic matrix;

[0007] Step S2: Perform multi-level clustering on the multi-dimensional operational feature matrix, and obtain typical conversion scenarios and their thermal stress characteristic parameters by screening representative indicators of the scenarios;

[0008] Step S3: Evaluate the thermal performance of multiple alternative topologies under the typical conversion scenario and construct a scenario-topology adaptation matrix;

[0009] Step S4: Predict the cumulative thermal stress of devices in the future time period under the current topology, query the scenario-topology adaptation matrix to obtain the cumulative thermal stress of candidate topologies, take the difference between the two as the switching benefit, calculate the energy loss of device turn-off transient, inrush current and restart process during topology switching and convert it into equivalent temperature rise as the switching cost, and perform topology switching when the net benefit of switching benefit minus the switching cost is greater than the set threshold in three consecutive rolling prediction cycles.

[0010] Secondly, this application provides a power converter optimization configuration system for power distribution networks based on scenario clustering, the power converter optimization configuration system for power distribution networks based on scenario clustering includes:

[0011] The extraction module is used to collect the input voltage, output current, instantaneous conversion power, and junction temperature of each switching device of the power converter, calculate the power change rate and cumulative thermal stress index, and form a multi-dimensional operating characteristic matrix.

[0012] The clustering module is used to perform multi-level clustering on the multi-dimensional operating feature matrix, and to obtain typical conversion scenarios and their thermal stress characteristic parameters by screening representative indicators of the scenarios.

[0013] The evaluation module is used to evaluate the thermal performance of multiple alternative topologies under the typical conversion scenario and construct a scenario-topology adaptation matrix.

[0014] The query module is used to predict the cumulative thermal stress of devices in the future time period under the current topology, query the scenario-topology adaptation matrix to obtain the cumulative thermal stress of candidate topologies, and use the difference between the two as the switching benefit. The energy loss of device turn-off transient, inrush current and restart process during the topology switching process is converted into equivalent temperature rise as the switching cost. When the net benefit of the switching benefit minus the switching cost is greater than a set threshold in three consecutive rolling prediction cycles, the topology switching is performed.

[0015] Thirdly, a power converter optimization configuration device based on scenario clustering is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the power converter optimization configuration device based on scenario clustering to execute the aforementioned power converter optimization configuration method based on scenario clustering.

[0016] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to execute the above-described method for optimizing the configuration of power converters in a power distribution network based on scenario clustering.

[0017] The technical solution provided in this application establishes a three-domain coupled data acquisition system for power conversion equipment, which simultaneously acquires multi-dimensional operating parameters such as input voltage, output current, instantaneous conversion power, and junction temperature of each switching device. It also calculates the power change rate to reflect the severity of power fluctuations and the cumulative thermal stress index to reflect the nonlinear cumulative effect of thermal damage to the device. The resulting multi-dimensional operating feature matrix breaks through the limitation of existing technologies that only collect macroscopic parameters such as power, and provides a comprehensive data foundation for subsequent scenario clustering, including transient fluctuation characteristics, thermal accumulation dynamics, and device temperature distribution. By performing multi-level clustering on the multi-dimensional operational feature matrix, K-means clustering is used to identify transient power patterns for rapidly changing features such as power fluctuation intensity, hierarchical clustering is used to distinguish heat load states for slowly changing features such as accumulated thermal stress indicators, and DBSCAN density clustering is used to extract typical daily patterns for diurnal changing features such as load rate. This multi-level clustering architecture matches the multi-timescale characteristics of distribution network operation scenarios. Compared with the single-level clustering of existing technologies, it can systematically separate scenario features in three dimensions: rapid fluctuation, heat accumulation, and diurnal cycle, avoiding clustering distortion caused by the mixing of features from different time scales. The scenario representativeness index is obtained by weighted summation of the sample number ratio, thermal damage ratio, and clustering quality. While screening high-frequency scenarios, extreme samples with extreme temperatures, rapid heating, and uneven temperature are forcibly retained. This ensures that the typical transformation scenario library covers both common operating modes and rare but serious extreme scenarios, providing comprehensive scenario coverage for the construction of the topology adaptation matrix. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of an embodiment of the power converter optimization configuration method for power distribution networks based on scenario clustering in this application.

[0020] Figure 2 This is a schematic diagram of an embodiment of the power converter optimization configuration system for power distribution networks based on scenario clustering in this application.

[0021] Figure 3 This is a schematic block diagram of the structure of the power converter optimization configuration device for power distribution networks based on scenario clustering in an embodiment of the present invention. Detailed Implementation

[0022] This application provides a method and system for optimizing the configuration of power converters in a power distribution network based on scenario clustering. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0023] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the power converter optimization configuration method for power distribution networks based on scenario clustering in this application includes:

[0024] Step S1: Collect the input voltage, output current, instantaneous conversion power, and junction temperature of each switching device of the power converter, calculate the power change rate and cumulative thermal stress index, and construct a multi-dimensional operating characteristic matrix;

[0025] Specifically, the input voltage acquisition includes instantaneous values ​​and phase angle information of the three-phase voltage. The output current records not only the amplitude but also its time derivative. The junction temperature of each switching device is directly measured in real time by the NTC thermistor built into the IGBT module. The power change rate is calculated by subtracting the power value at the previous sampling moment from the current power value and then dividing by the sampling time interval, reflecting the severity of power fluctuations. The cumulative thermal stress index uses a time integration method, accumulating the portion of the junction temperature exceeding the 85℃ reference temperature by power 1.5 of the temperature minus 85℃. This exponential form is based on the Coffin-Manson fatigue model, reflecting the nonlinear accelerating effect of temperature on device lifespan damage. The multidimensional operating characteristic matrix normalizes electrical parameters to the 0-1 interval using minimum-maximum value normalization to eliminate dimensional differences. Temperature parameters are standardized by subtracting the mean and dividing by the standard deviation to ensure the data distribution conforms to a standard normal distribution.

[0026] Step S2: Perform multi-level clustering on the multi-dimensional operational feature matrix, and obtain typical conversion scenarios and their thermal stress characteristic parameters by selecting representative indicators of the scenarios;

[0027] Specifically, multi-level clustering addresses the multi-timescale characteristics of power conversion scenarios in distribution networks. The first level extracts power fluctuation intensity, current change rate, and junction temperature rise rate to form a fast-changing feature set, which is then used for K-means clustering. The optimal cluster number is determined by calculating the silhouette coefficient and elbow rule curve under different cluster numbers, identifying transient power patterns such as step changes, oscillations, and slow rises. The second level uses accumulated thermal stress indicators, average junction temperature within past time windows, inter-device temperature standard deviation, and coolant temperature to form a slow-changing feature set, performing hierarchical clustering based on the Ward connection criterion. Cluster granularity is controlled by setting an intra-cluster variance threshold, distinguishing between thermal load states such as low-temperature light load, medium-temperature medium load, high-temperature heavy load, and insufficient heat dissipation. The third level performs DBSCAN density clustering based on load factor, power factor, ambient temperature, and time-based sine and cosine encoding. This algorithm automatically identifies high-density regions as cluster centers using two parameters: neighborhood radius and minimum sample size. It can discover daily operating pattern clusters of arbitrary shapes and mark low-density regions as noise points. The representativeness indicators for a scenario include: the proportion of the number of samples to the total number of samples, reflecting the frequency of the scenario; the ratio of the average cumulative thermal damage within the scenario to the global maximum thermal damage, reflecting the degree of thermal threat; and the reciprocal of the ratio of the standard deviation within a cluster to the distance from the centroid of that cluster to the centroid of the nearest neighbor cluster, reflecting the cluster compactness. Scenarios exceeding a threshold score are selected after a weighted sum of these three indicators. Extreme sample identification sets three threshold conditions: junction temperature exceeding 130℃, junction temperature rise rate exceeding 5℃ per minute, and inter-device temperature deviation exceeding 22℃. Samples meeting any of these conditions are extracted and then sub-clustered to form extreme scenario modes such as startup shock, continuous overload, and heat dissipation failure, which are then forcibly included in the typical scenario library.

[0028] Step S3: Evaluate the thermal performance of multiple alternative topologies under typical conversion scenarios and construct a scenario-topology adaptation matrix;

[0029] Specifically, for the 10kW-level power conversion requirement of DC 600V to AC 380V, various topologies were designed, including two-level full-bridge, three-level NPC, interleaved parallel, T-type three-level, and Vienna rectifier cascade. For each topology, a refined simulation system was built on the PLECS simulation platform, including an IGBT switching characteristic lookup table model and a fourth-order Cauer thermal network model. The IGBT switching characteristic model established a three-dimensional interpolation table of turn-on energy, turn-off energy, and on-state voltage drop with respect to current, voltage, and junction temperature based on the device datasheet. The fourth-order thermal network decomposed the thermal resistance from the junction to the housing into four stages of series resistor-capacitor networks, where the thermal resistance and thermal capacitance values ​​of each stage were obtained by fitting the transient thermal impedance curve provided by the manufacturer. The characteristic centroid vectors of typical conversion scenarios are converted into operating parameters such as power, output current, power factor, input voltage, and ambient temperature, which are then substituted into the simulation models of each topology. The initial junction temperature is set to the ambient temperature plus 10°C, and the simulation is run for 120 minutes. The first 60 minutes are the hot start transition phase, and the next 60 minutes are the steady state phase. During the steady state phase, the junction temperature values ​​of each device are sampled every 10 seconds to extract thermal performance indicators such as the maximum junction temperature peak, time average, and inter-device variance. At the same time, the input power and output power are recorded to calculate the conversion efficiency, and the output voltage waveform is calculated using fast Fourier transform to calculate the total harmonic distortion rate and other electrical performance indicators. The multi-objective comprehensive scoring function assigns a weight of 0.35 to the maximum junction temperature peak value after normalization by subtracting the actual peak value from 150℃, a weight of 0.25 to the conversion efficiency after direct normalization, a weight of 0.2 to the inter-device temperature variance after normalization by taking the reciprocal, a weight of 0.1 to the output voltage harmonic distortion rate after normalization by taking the reciprocal, and a weight of 0.1 to the safety margin value after normalization by subtracting the maximum junction temperature from the 125℃ safety threshold. The weighted sum of these five values ​​yields the comprehensive score for each topology in each scenario, which is then filled into the scenario-topology adaptation matrix.

[0030] Step S4: Predict the cumulative thermal stress of devices in the future time period under the current topology, query the scenario-topology adaptation matrix to obtain the cumulative thermal stress of candidate topologies, and use the difference between the two as the switching benefit. Calculate the energy loss of device turn-off transient, inrush current and restart process during topology switching and convert it into equivalent temperature rise as the switching cost. When the net benefit of switching benefit minus switching cost is greater than the set threshold in three consecutive rolling prediction cycles, the topology switch is executed.

[0031] Specifically, the constructed LSTM-CNN hybrid neural network receives a 12-dimensional feature sequence containing 120 time steps over the past 60 minutes as input. The forward layer of the bidirectional LSTM propagates from the start point to the end point to capture the cumulative effect of historical thermal stress, while the backward layer propagates from the end point to the start point to capture the constraints of future boundary conditions on the current state. The hidden state vectors from both directions are concatenated to form a 256-dimensional temporal feature vector. The CNN branch sets three parallel convolutional paths with kernel sizes of 5, 11, and 21 to extract local patterns over time spans of 2.5 minutes, 5.5 minutes, and 10.5 minutes, respectively. The short-term convolutional kernel captures the transient temperature response caused by rapid power fluctuations, the medium-term convolutional kernel captures the thermal accumulation inertial features, and the long-term convolutional kernel captures the hysteresis characteristics of the heat dissipation system. The outputs of the three convolutional paths are concatenated after max pooling and global average pooling to form a 256-dimensional spatial feature vector. The attention fusion layer calculates dynamic weight coefficients for temporal and spatial feature vectors. These weights are obtained by multiplying the trainable parameter matrix with the concatenated feature vectors and then normalizing using the softmax function, achieving adaptive weighting for long-term LSTM dependencies and local CNN patterns. The fully connected prediction layer outputs a junction temperature prediction matrix multiplied by 60 time steps and 6 device parameters. Based on this matrix, the maximum junction temperature value at each time step is extracted, and the portion exceeding the 85°C reference temperature is raised to the power of 1.5 and integrated over time using the trapezoidal rule to obtain the predicted cumulative thermal damage. When querying the scenario-topology adaptation matrix, the predicted future average power, average output current, and average power factor are first combined to form a prediction parameter vector. The Mahalanobis distance between this vector and the centroid vectors of each typical scenario is calculated. The Mahalanobis distance calculation formula is: the predicted vector minus the transpose of the centroid vector, multiplied by the inverse of the scenario's covariance matrix, multiplied by the square root of the predicted vector minus the centroid vector. The scenario with the smallest Mahalanobis distance is selected as the matching scenario. The maximum junction temperature peak value corresponding to each candidate topology in the adaptation matrix is ​​extracted. The peak value is subtracted from 85℃, raised to the power of 1.5, and then multiplied by 30 minutes to estimate the cumulative thermal stress of the candidate topology. The switching cost calculation includes: the square of the rated voltage of the bus capacitor multiplied by the capacitance value multiplied by 0.19 to obtain the discharge energy during the turn-off stage; the square of the impact charging current of the output filter capacitor multiplied by the equivalent series resistance multiplied by the charging duration to obtain the energy during the dead zone stage; the average power multiplied by 0.06 to obtain the additional loss energy during the soft-start stage; and the sum of the three energies divided by the device heat capacity multiplied by the number of devices to obtain the equivalent temperature rise.The rolling prediction mechanism does not make an immediate decision after obtaining a preliminary prediction at the current moment. It delays for 10 time steps until the first delay moment, when the measured data for that period has been obtained. It then makes a new prediction using a new time window containing the measured data. After another 10 time steps, it makes a third prediction until the second delay moment. The switching is only performed when all three predictions determine that the same alternative topology should be switched to, and the net gain of the predicted thermal damage of the current topology minus the estimated thermal damage of the alternative topology minus twice the equivalent temperature rise is greater than the set threshold of 10°C raised to the power of 1.5 multiplied by minutes. During the switching process, in the power reduction phase, the PWM modulation is reduced from the current value to 0.2 at a slope of 0.25 per second, reducing the output current to 15% of the rated value in 3 seconds. In the safety disconnect phase, after blocking all gate drive signals, it waits for 10 milliseconds to ensure that the device is completely turned off. In the topology reconstruction phase, the PWM channel phase relationship and switching frequency parameters are configured according to the new topology type. In the soft start phase, the modulation is linearly increased to the target value at a slope of 0.16 per second in 5 seconds.

[0032] In one specific embodiment, step S1 includes:

[0033] The power converter's operating data is continuously collected at a set sampling interval. The operating data includes input voltage, output current, instantaneous conversion power, power factor, junction temperature of each switching device, heat sink temperature, ambient temperature, and cooling fan speed.

[0034] Remove outliers from the operating data where the junction temperature of each switching device exceeds the upper limit threshold or is lower than the set difference of the ambient temperature, and use interpolation to fill in the missing data.

[0035] Based on the operating data, the power change rate, junction temperature change rate, cumulative thermal stress index and device temperature fluctuation amplitude are calculated. The cumulative thermal stress index is the temperature-time accumulation of each switching device when the junction temperature is above the set reference temperature.

[0036] The input voltage, output current, and instantaneous conversion power are normalized, and the junction temperature of each switching device is standardized to form a multi-dimensional operating characteristic matrix.

[0037] Specifically, a continuous data acquisition mechanism for the power converter in the distribution network is established with a sampling interval of 30 seconds. When acquiring the input voltage, the instantaneous amplitude and phase angle of the three-phase voltage are recorded. The output current includes the current amplitude and harmonic distortion rate. The instantaneous conversion power is calculated by summing the products of the three-phase voltage and the corresponding phase current and then dividing by 3. The power factor is obtained by the ratio of active power to apparent power. The junction temperature of each switching device is directly measured by the NTC thermistor embedded in the IGBT module to obtain the real-time temperature of the six main switching transistors. The heat sink temperature is measured non-contactly using an infrared temperature sensor. The ambient temperature and cooling fan speed are acquired by a temperature and humidity sensor and a Hall speed sensor, respectively. When removing outliers, the junction temperature of each switching device is compared with the upper limit threshold of 150℃. If it exceeds this value, it is determined to be a sensor fault. At the same time, it is checked whether the junction temperature is lower than the ambient temperature minus 3℃. If it is lower than this threshold, it violates the second law of thermodynamics and is also determined to be an anomaly. The isolated forest algorithm is used to identify outliers in the multidimensional feature space. For missing data, cubic spline interpolation is used to fit a cubic polynomial curve to complete the missing data by using data from 10 sampling times before and after the missing point.

[0038] The power change rate is calculated by subtracting the power value from the previous sampling time from the current power value and then dividing by a 30-second sampling interval, in watts per second. The junction temperature change rate is calculated using the same method: subtracting the previous junction temperature from the current junction temperature and dividing by the time interval, in degrees Celsius per second. The cumulative thermal stress index uses a time-integral accumulation method. At each sampling time, the maximum junction temperature value of each switching device is extracted. When this value exceeds the 85°C reference temperature, the temperature difference exceeding this value is calculated. This temperature difference raised to the power of 1.5 is multiplied by a 30-second time step and accumulated to the historical cumulative value. This 1.5-power exponent reflects the nonlinear accelerating effect of temperature on IGBT bond wire and solder layer fatigue damage based on the Coffin-Manson fatigue accumulation model. The device temperature fluctuation range is obtained by extracting the junction temperature sequence of 120 sampling points over the past 60 minutes and calculating the maximum value minus the minimum value of this sequence. Normalization is applied to electrical parameters such as input voltage, output current, and instantaneous conversion power. A minimum-maximum normalization method is used, subtracting the minimum value from all samples and then dividing by the difference between the maximum and minimum values, mapping the current value to the 0-1 interval to eliminate dimensional differences. Standardization is applied to the junction temperature of each switching device, using the Z-score method to subtract the mean from all samples and then divide by the standard deviation, ensuring the temperature data conforms to a standard normal distribution with a mean of 0 and a standard deviation of 1. This method is more robust to abnormally high temperatures. Finally, the normalized and standardized 58-dimensional feature vectors are arranged in chronological order to form a multi-dimensional operational feature matrix.

[0039] In one specific embodiment, step S2 includes:

[0040] Feature filtering is performed on the multidimensional running feature matrix based on mutual information coefficients and feature importance to obtain a dimensionality-reduced feature matrix;

[0041] K-means clustering is performed on the fast-changing features in the dimensionality-reduced feature matrix to obtain the power transient mode, hierarchical clustering is performed on the slow-changing features to obtain the heat load state, and DBSCAN density clustering is performed on the diurnal-changing features to obtain the typical daily mode. These are combined to form a preliminary scene category.

[0042] Calculate and screen representative indicators for each preliminary scenario category. The representative indicators are obtained by weighted summation of the sample size ratio, thermal damage ratio, and cluster quality.

[0043] Identify extreme samples with extreme temperatures, rapid heating, and uneven temperature distribution. After sub-clustering various extreme samples, force their inclusion to form typical transition scenarios and their characteristic centroid vectors, thermal stress characteristic parameters, and scenario transition probabilities.

[0044] Specifically, the mutual information coefficient uses the maximum mutual information coefficient method to measure the nonlinear correlation between each feature parameter and the maximum junction temperature of key components. This is achieved by constructing grid partitions of different resolutions in a two-dimensional feature space, calculating the mutual information value under each partition, and taking the maximum value, with the value ranging from 0 to 1. Feature importance is determined by constructing an XGBoost gradient boosting regression model to predict the maximum junction temperature of key components. The model contains 200 decision trees with a maximum depth of 8 layers and a learning rate of 0.05. After training, the gain importance of each feature is extracted. This metric measures the average loss function reduction brought by the feature at all decision tree split nodes. Features with a mutual information coefficient greater than 0.6 and ranking in the top 20 in gain importance are selected. Finally, 12 core features are selected to form a dimensionality-reduced feature matrix: power, output current, power factor, cumulative thermal damage, junction temperature rise rate, inter-device temperature deviation, power fluctuation intensity, load factor, switching frequency, coolant temperature, current change rate, and current harmonic distortion rate. Rapidly changing characteristics include power fluctuation intensity, current change rate, and junction temperature rise rate. The K-means clustering algorithm is used to minimize the sum of squared Euclidean distances from samples within a cluster to the centroid through iterative optimization. The K-means++ initialization method is used to select initial centroids that are far apart from each other. The clustering effect is evaluated by calculating the Calinski-Harabasz index, which is the ratio of inter-cluster variance to intra-cluster variance. Finally, 12 clusters are determined to identify step, oscillatory, and slowly changing power transient modes.

[0045] Slow-varying features include cumulative thermal damage, average junction temperature, inter-device temperature deviation, and coolant temperature. The Ward connection criterion for hierarchical clustering is employed, which selects the two clusters with the smallest increase in intra-cluster squared error during each cluster merging. The quality of the hierarchical structure is evaluated by calculating the cophenetic correlation coefficient of the dendrogram. A cutting threshold is set to ensure that the intra-cluster variance does not exceed 0.18, resulting in seven clusters that distinguish thermal load states such as low temperature light load, medium temperature medium load, high temperature heavy load, and insufficient heat dissipation. Diurnal variation features include load factor, power factor, ambient temperature, and sine and cosine encoding of time. DBSCAN density clustering is used, defining core points with a neighborhood radius of 0.28 and a minimum sample size of 80. Points with density reachable between core points are grouped into the same cluster, and low-density areas are marked as noise points, resulting in six clusters that identify typical daily patterns such as weekday peaks, nighttime troughs, high-temperature days, and low-temperature days. The representative indicators include the following: the sample number ratio is calculated by dividing the sample number by the total sample number and then multiplying by the natural logarithm of the duration of the scene plus 1; the thermal damage ratio is calculated by dividing the average cumulative thermal damage within the scene by the global maximum thermal damage value and then raising it to the power of 1.2; and the clustering quality is calculated by subtracting the standard deviation within the cluster from 1 and dividing by the Euclidean distance from the centroid of the cluster to the centroid of the nearest neighbor cluster. The three items are weighted and summed with weight coefficients of 0.35, 0.45, and 0.20, respectively. When identifying extreme temperature samples, the highest junction temperature was set to exceed 130℃; when identifying rapidly heating samples, the junction temperature rise rate was set to exceed 5℃ per minute; and when identifying temperature unevenness samples, the temperature deviation between devices was set to exceed 22℃. After extracting samples that met each threshold condition, the K-means algorithm was used for sub-clustering. Extreme temperature samples were clustered into 4 classes, rapidly heating samples into 3 classes, and temperature unevenness samples into 3 classes, for a total of 10 extreme scene modes, which were forcibly included into the typical scene library, ultimately forming 78 typical transition scenes. The mean vector of all samples in the 12-dimensional feature space of each scene was calculated as the feature centroid vector, and the feature covariance matrix was calculated for subsequent Mahalanobis distance scene recognition. The scene transition probability matrix was obtained by dividing the number of transitions between each scene by the total number of occurrences of that scene based on time series statistics.

[0046] In one specific embodiment, step S3 includes:

[0047] Multiple alternative topologies are designed for the same power conversion function, and IGBT switching characteristic models and fourth-order thermal network models of each alternative topology are established in the simulation platform.

[0048] Substitute the running parameters corresponding to the feature centroid vectors of each typical conversion scenario into the simulation model of each candidate topology, and after running for a set time, extract the maximum junction temperature peak, average junction temperature, inter-device temperature variance, total power loss and conversion efficiency in the steady state stage.

[0049] A multi-objective comprehensive scoring function is constructed based on the maximum junction temperature peak, conversion efficiency, inter-device temperature variance, and output voltage harmonic distortion rate to calculate the comprehensive score of each candidate topology in each typical conversion scenario.

[0050] A scenario-topology adaptation matrix is ​​constructed using typical conversion scenarios as rows, alternative topologies as columns, and comprehensive scores as matrix elements.

[0051] Specifically, the mutual information coefficient uses the maximum mutual information coefficient method to measure the nonlinear correlation between each feature parameter and the highest junction temperature of key components. The mutual information value is calculated by constructing grids of different resolutions in a two-dimensional feature space and taking the maximum value, with the value ranging from 0 to 1. Feature importance is used to predict the highest junction temperature of key components using an XGBoost gradient boosting regression model. The model contains 200 decision trees with a maximum depth of 8 layers and a learning rate of 0.05. The gain importance of each feature is extracted; this metric measures the decrease in the average loss function brought by the feature at the split nodes of the decision trees. Features with a mutual information coefficient greater than 0.6 and ranking in the top 20 in gain importance are selected. Twelve core features—power, output current, power factor, cumulative thermal damage, junction temperature rise rate, inter-device temperature deviation, power fluctuation intensity, load factor, switching frequency, coolant temperature, current change rate, and current harmonic distortion rate—are used to construct a dimensionality-reduced feature matrix. Rapidly changing characteristics include power fluctuation intensity, current change rate, and junction temperature rise rate. The K-means clustering algorithm is used to minimize the sum of squared Euclidean distances from samples within a cluster to the centroid through iterative optimization. The K-means++ initialization method is used to select initial centroids that are farther away. The clustering effect is evaluated by the Calinski-Harabasz index, which is the ratio of inter-cluster variance to intra-cluster variance. Finally, 12 clusters are determined to identify step, oscillatory, and slowly changing power transient modes.

[0052] Slow-varying features include cumulative thermal damage, average junction temperature, inter-device temperature deviation, and coolant temperature. The Ward connection criterion for hierarchical clustering is employed, which merges the two clusters with the smallest increase in intra-cluster squared error each time. The quality of the hierarchical structure is evaluated using the cophenetic correlation coefficient of the dendrogram. A cutting threshold is set to ensure that the intra-cluster variance does not exceed 0.18, resulting in 7 clusters that distinguish thermal load states such as low temperature light load, medium temperature medium load, high temperature heavy load, and insufficient heat dissipation. Diurnal variation features include load rate, power factor, ambient temperature, and sine and cosine encoding of time. DBSCAN density clustering is used, defining core points with a neighborhood radius of 0.28 and a minimum sample size of 80. Points with reachable density are grouped into the same cluster, and low-density areas are marked as noise points, resulting in 6 clusters that identify typical daily patterns such as weekday peaks, nighttime troughs, high-temperature days, and low-temperature days. Among the representative indicators, the sample number ratio is calculated by dividing the number of scene samples by the total number of samples and then multiplying by the natural logarithm of the scene duration ratio plus 1; the thermal damage ratio is calculated by dividing the average cumulative thermal damage within the scene by the 1.2 power of the global maximum thermal damage value; and the clustering quality is calculated by subtracting the standard deviation within the cluster from 1 and dividing by the Euclidean distance from the centroid of the cluster to the centroid of the nearest neighbor cluster. The three indicators are weighted and summed with weight coefficients of 0.35, 0.45, and 0.20, respectively. Extreme temperature samples were identified by setting a maximum junction temperature exceeding 130℃; rapid temperature rise samples were identified by setting a junction temperature rise rate exceeding 5℃ per minute; and temperature unevenness samples were identified by setting a temperature deviation between devices exceeding 22℃. After extracting samples that met each threshold condition, K-means sub-clustering was used. Extreme temperature samples were clustered into 4 classes, rapid temperature rise samples into 3 classes, and temperature unevenness samples into 3 classes, for a total of 10 extreme scenario modes, which were forcibly incorporated to form 78 typical transition scenarios. The mean vector of all samples in the 12-dimensional feature space of each scenario was calculated as the feature centroid vector. The feature covariance matrix was calculated for Mahalanobis distance scene recognition. The number of transitions between each scenario was counted and divided by the total number of occurrences of that scenario to obtain the scene transition probability matrix.

[0053] In one specific embodiment, step S4, predicting the cumulative thermal stress of the device over a future time period under the current topology, includes:

[0054] Construct an LSTM-CNN hybrid neural network model, with the input being a multidimensional feature sequence containing a sliding time window of a set past duration;

[0055] The LSTM-CNN hybrid neural network model includes a two-layer bidirectional LSTM branch, a multi-scale parallel convolutional CNN branch, an attention fusion layer, and a fully connected prediction layer. The forward and backward hidden states of the two-layer bidirectional LSTM branch are concatenated to obtain the temporal feature vector. The multi-scale parallel convolutional CNN branch includes short-term convolutional paths, mid-term convolutional paths, and long-term convolutional paths. The outputs of the three convolutional paths are pooled and concatenated to obtain the spatial feature vector.

[0056] The attention fusion layer calculates the dynamic weights of the temporal feature vector and the spatial feature vector and fuses them in a weighted manner. The fully connected prediction layer outputs the multi-device junction temperature prediction matrix for multiple future time steps.

[0057] The predicted peak junction temperature and the predicted cumulative thermal damage are calculated based on the multi-device junction temperature prediction matrix. The predicted cumulative thermal damage is the integral of the predicted junction temperature raised to a set power over time.

[0058] Specifically, the input to the LSTM-CNN hybrid neural network model is a multi-dimensional feature sequence containing a sliding time window of the past 60 minutes. Specifically, it is a two-dimensional matrix of 120 time steps multiplied by 12-dimensional features. Each time step corresponds to a 30-second sampling interval. The 12-dimensional features include power, output current, power factor, cumulative thermal damage, junction temperature rise rate, inter-device temperature deviation, power fluctuation intensity, load factor, switching frequency, coolant temperature, current change rate, and current harmonic distortion rate. The first layer of the two-layer bidirectional LSTM branch contains 256 hidden units each in the forward LSTM and backward LSTM. The forward LSTM propagates from the starting point to the ending point to capture the cumulative effect of historical thermal stress, while the backward LSTM propagates from the ending point to the starting point to capture the constraints of future boundary conditions on the current state. The hidden state vectors of the two directions at each time step are concatenated to form a 512-dimensional vector. The second layer also adopts a bidirectional structure with 128 hidden units each. Finally, the forward and backward hidden states are concatenated to obtain a 256-dimensional temporal feature vector, which encodes the long-term dependence of thermal stress evolution over the past 60 minutes. The multi-scale parallel convolutional CNN branch treats the 120x12 input feature matrix as a two-dimensional feature map. The short-term convolution path uses a kernel size of 5 corresponding to a 2.5-minute time span, containing 64 convolution kernels to extract the coupling mode of rapid power fluctuations and transient temperature response along the time dimension. The medium-term convolution path uses a kernel size of 11 corresponding to a 5.5-minute time span, containing 128 convolution kernels to extract thermal accumulation inertia features. The long-term convolution path uses a kernel size of 21 corresponding to a 10.5-minute time span, containing 64 convolution kernels to extract the response hysteresis characteristics of the heat dissipation system.

[0059] The three convolutional outputs are each passed through a max pooling layer with a pooling window size of 2 to reduce the feature map size. Then, they are concatenated along the channel dimension to form a 256-dimensional feature vector. This vector is then averaged along the temporal dimension by a global average pooling layer to obtain a 256-dimensional spatial feature vector. This vector encodes the power-current-temperature spatial coupling relationship across multiple time scales. The attention fusion layer first concatenates the 256-dimensional temporal and spatial feature vectors into a 512-dimensional vector. A trainable weight matrix maps this 512-dimensional vector to a 2-dimensional attention score vector. A softmax function is applied to this score vector to normalize it, obtaining dynamic weight coefficients for the temporal and spatial features. The sum of these two weight coefficients is 1. Multiplying the temporal feature vector by its corresponding weight coefficient and adding it to the spatial feature vector multiplied by its corresponding weight coefficient yields the 256-dimensional fused feature vector. The fully connected prediction layer consists of two fully connected sub-layers. The first sub-layer maps the 256-dimensional fused feature vector to a 512-dimensional vector and applies a ReLU activation function with random deactivation at a dropout ratio of 0.3. The second sub-layer maps the 512-dimensional vector to a 360-dimensional vector. This 360-dimensional vector is reconstructed into a 2D matrix multiplied by 6 devices over 60 time steps, serving as the multi-device junction temperature prediction matrix. Each element of the matrix represents the predicted junction temperature value of a device at a future time step. The row index represents the next 30 minutes for time steps 1 to 60, and the column index represents the 6 main switching devices. When calculating the predicted peak junction temperature based on the multi-device junction temperature prediction matrix, the maximum value among the 6 devices is extracted from each row of the matrix to form a sequence of the highest junction temperatures over 60 time steps. The maximum value of this sequence is the predicted peak junction temperature. The calculation of the predicted cumulative thermal damage first involves subtracting the 85°C reference temperature from each element in the highest junction temperature sequence over 60 time steps to obtain the overtemperature sequence. Elements in the overtemperature sequence that are less than 0 are set to 0. The damage intensity sequence is obtained by raising the non-negative element to the power of 1.5. This sequence is then integrated over the time axis using the trapezoidal rule, which is the average damage intensity of two adjacent time steps multiplied by a 30-second time step. The integral value of 60 time steps is accumulated to obtain the predicted cumulative thermal damage, expressed in degrees Celsius raised to the power of 1.5 multiplied by minutes.

[0060] In one specific embodiment, step S4, querying the scene-topology adaptation matrix to obtain the accumulated thermal stress of the candidate topology, includes:

[0061] The average power, average output current, and average power factor for future time periods are calculated based on the multi-device junction temperature prediction matrix.

[0062] The Mahalanobis distance between the predicted parameter vector composed of average power, average output current and average power factor and the feature centroid vector of each typical conversion scenario is calculated, and the scenario with the smallest Mahalanobis distance is identified as the matching scenario.

[0063] In the scenario-topology adaptation matrix, query the maximum junction temperature peak value of each candidate topology corresponding to the matching scenario, and estimate the cumulative thermal stress of each candidate topology based on the maximum junction temperature peak value.

[0064] The difference between the predicted cumulative thermal damage of the current topology and the cumulative thermal stress of each alternative topology is calculated as the switching benefit of switching to each alternative topology.

[0065] Specifically, the input to the LSTM-CNN hybrid neural network model is a two-dimensional matrix consisting of 120 time steps multiplied by 12-dimensional features. Each time step corresponds to a 30-second sampling interval covering the past 60 minutes of running data. The 12-dimensional features include power, output current, power factor, cumulative thermal damage, junction temperature rise rate, inter-device temperature deviation, power fluctuation intensity, load factor, switching frequency, coolant temperature, current change rate, and current harmonic distortion rate. The first layer of the two-layer bidirectional LSTM branch contains 256 hidden units each for the forward LSTM and the backward LSTM. The forward LSTM propagates from the starting point to the ending point to capture the cumulative effect of historical thermal stress, while the backward LSTM propagates from the ending point to the starting point to capture the constraints of future boundary conditions on the current state. The hidden state vectors in the two directions are concatenated to form a 512-dimensional vector. The second bidirectional structure contains 128 hidden units each. Finally, the output forward and backward hidden states are concatenated to obtain a 256-dimensional temporal feature vector.

[0066] The multi-scale parallel convolutional CNN branch treats the 120x12 input matrix as a two-dimensional feature map. The short-term convolution path uses a kernel size of 5 containing 64 kernels to extract the coupling mode of rapid power fluctuation and transient temperature response over a 2.5-minute time span. The medium-term convolution path uses a kernel size of 11 containing 128 kernels to extract the thermal accumulation inertia feature over a 5.5-minute time span. The long-term convolution path uses a kernel size of 21 containing 64 kernels to extract the heat dissipation system response hysteresis characteristics over a 10.5-minute time span. The outputs of the three convolutions are concatenated in the channel dimension after passing through a max pooling layer with a pooling window size of 2, and then averaged in the time dimension by a global average pooling layer to obtain a 256-dimensional spatial feature vector. The attention fusion layer concatenates the 256-dimensional temporal feature vector and the 256-dimensional spatial feature vector into a 512-dimensional vector. This vector is then mapped to a 2-dimensional attention score vector using a trainable weight matrix. The score vector is normalized by applying a softmax function to obtain the dynamic weight coefficients of the temporal and spatial features. The sum of the two weight coefficients is 1. The temporal feature vector is multiplied by its weight coefficient, and the spatial feature vector is multiplied by its weight coefficient to obtain the 256-dimensional fused feature vector.

[0067] The fully connected prediction layer consists of two fully connected sub-layers. The first sub-layer maps the 256-dimensional fused feature vector to a 512-dimensional vector and applies a ReLU activation function and a random deactivation dropout with a dropout ratio of 0.3. The second sub-layer maps the 512-dimensional vector to a 360-dimensional vector. This vector is reconstructed into a 2D matrix multiplied by 6 devices over 60 time steps as a multi-device junction temperature prediction matrix. The row index represents the next 30 minutes for time steps 1 to 60, and the column index represents the 6 main switching devices. Based on this matrix, 6 devices are extracted from each row. The maximum value in the data forms a sequence of the highest junction temperatures over 60 time steps. The maximum value in this sequence is the predicted peak junction temperature. The calculation of the predicted cumulative thermal damage first involves subtracting the 85°C reference temperature from each element in the highest junction temperature sequence to obtain the overheat sequence. Elements in the overheat sequence that are less than 0 are set to 0. The damage intensity sequence is obtained by raising the non-negative elements to the power of 1.5. This sequence is then integrated using the trapezoidal rule, which is the average damage intensity of two adjacent time steps multiplied by a 30-second time step. The integral value of 60 time steps is accumulated to obtain the predicted cumulative thermal damage.

[0068] In one specific embodiment, step S4, calculating the switching cost of the topology handover process, includes:

[0069] The total switching energy cost is calculated by summing the energy discharged by the bus capacitor during the shutdown phase, the energy charged by the output filter capacitor during the dead zone phase, and the additional energy loss due to efficiency reduction during the soft start phase.

[0070] Divide the total switching energy cost by the product of the device's heat capacity and the number of devices to obtain the equivalent temperature rise as the switching cost.

[0071] A rolling prediction mechanism is adopted, and predictions are made at the current time, the first delay time after the current time, and the second delay time after the current time. When all three predictions determine that the switch to the same alternative topology and the net benefit is greater than the set threshold, the switching process of power reduction, safe disconnection, topology reconstruction, soft start and stability confirmation is executed.

[0072] During the switching process, the power reduction phase linearly reduces the PWM modulation to the set ratio, the safety disconnect phase blocks the gate drive and waits for the dead time, the topology reconfiguration phase configures the PWM channel mapping and switching frequency of the new topology, and the soft start phase linearly increases the modulation to the target value.

[0073] Specifically, the total switching energy cost includes the energy discharged by the bus capacitor during the shutdown phase (calculated by multiplying the square of the bus voltage by the bus capacitance value by 0.19), the energy charged by the output filter capacitor during the dead-zone phase (calculated by multiplying the square of the inrush current by the equivalent series resistance of the filter capacitor by the charging duration), and the additional energy loss during the soft-start phase (calculated by multiplying the average power by 0.06). The sum of these three costs divided by the device's thermal capacity multiplied by the number of devices yields the equivalent temperature rise, which is the switching cost. The rolling prediction mechanism does not make an immediate decision after obtaining a preliminary prediction at the current moment. It delays for 10 time steps until the first delay time, at which point measured data for that period is obtained. A new time window containing the measured data is used for re-prediction. A third prediction is made after another 10 time steps until the second delay time. The switching process is executed when all three predictions determine that switching to the same alternative topology is necessary, and the net benefit of the predicted thermal damage of the current topology minus the estimated thermal damage of the alternative topology minus twice the equivalent temperature rise is greater than a set threshold. During the switching process, the power reduction phase linearly reduces the PWM modulation at a slope of 0.25 per second from the current value to 0.2, reducing the output current to 15% of the rated value in 3 seconds. In the safety disconnect phase, after blocking all gate drive signals, wait for 10 milliseconds to ensure that the device is completely turned off. In the topology reconfiguration phase, configure the PWM channel phase relationship and switching frequency parameters according to the new topology type. In the soft start phase, the modulation is linearly increased to the target value at a slope of 0.16 per second, completing the switching in 5 seconds.

[0074] The above describes the distribution network power converter optimization configuration method based on scenario clustering in the embodiments of this application. The following describes the distribution network power converter optimization configuration system based on scenario clustering in the embodiments of this application. One embodiment of the distribution network power converter optimization configuration system based on scenario clustering in the embodiments of this application includes:

[0075] The extraction module is used to collect the input voltage, output current, instantaneous conversion power, and junction temperature of each switching device of the power converter, calculate the power change rate and cumulative thermal stress index, and form a multi-dimensional operating characteristic matrix.

[0076] The clustering module is used to perform multi-level clustering on the multi-dimensional operating feature matrix, and to obtain typical conversion scenarios and their thermal stress characteristic parameters by screening representative indicators of the scenarios.

[0077] The evaluation module is used to evaluate the thermal performance of multiple alternative topologies under the typical conversion scenario and construct a scenario-topology adaptation matrix.

[0078] The query module is used to predict the cumulative thermal stress of devices in the future time period under the current topology, query the scenario-topology adaptation matrix to obtain the cumulative thermal stress of candidate topologies, and use the difference between the two as the switching benefit. The energy loss of device turn-off transient, inrush current and restart process during the topology switching process is converted into equivalent temperature rise as the switching cost. When the net benefit of the switching benefit minus the switching cost is greater than a set threshold in three consecutive rolling prediction cycles, the topology switching is performed.

[0079] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing the configuration of power converters in a power distribution network based on scenario clustering, characterized in that, The method includes: Step S1: Collect the input voltage, output current, instantaneous conversion power, and junction temperature of each switching device of the power converter, calculate the power change rate and cumulative thermal stress index, and construct a multi-dimensional operating characteristic matrix; Step S2: Perform multi-level clustering on the multi-dimensional operational feature matrix, and obtain typical conversion scenarios and their thermal stress characteristic parameters by screening representative indicators of the scenarios; Step S3: Evaluate the thermal performance of multiple alternative topologies under the typical conversion scenarios and construct a scenario-topology adaptation matrix. This includes: designing multiple alternative topologies for the same power conversion function; establishing IGBT switching characteristic models and fourth-order thermal network models for each alternative topology in a simulation platform; substituting the operating parameters corresponding to the feature centroid vectors of each typical conversion scenario into the simulation models of each alternative topology; extracting the maximum junction temperature peak, average junction temperature, inter-device temperature variance, total power loss, and conversion efficiency in the steady-state phase after running for a set duration; constructing a multi-objective comprehensive scoring function based on the maximum junction temperature peak, the conversion efficiency, the inter-device temperature variance, and the output voltage harmonic distortion rate; calculating the comprehensive score of each alternative topology under each typical conversion scenario; and constructing the scenario-topology adaptation matrix with the typical conversion scenarios as rows, the alternative topologies as columns, and the comprehensive score as matrix elements. Step S4: Predict the cumulative thermal stress of devices in the future time period under the current topology, query the scenario-topology adaptation matrix to obtain the cumulative thermal stress of candidate topologies, take the difference between the two as the switching benefit, calculate the energy loss of device turn-off transient, inrush current and restart process during topology switching and convert it into equivalent temperature rise as the switching cost, and perform topology switching when the net benefit of switching benefit minus the switching cost is greater than the set threshold in three consecutive rolling prediction cycles.

2. The method for optimizing the configuration of power converters in a distribution network based on scenario clustering according to claim 1, characterized in that, Step S1 includes: The power converter's operating data is continuously collected at a set sampling interval. The operating data includes the input voltage, the output current, the instantaneous conversion power, the power factor, the junction temperature of each switching device, the heat sink temperature, the ambient temperature, and the cooling fan speed. Remove outliers from the operating data where the junction temperature of each switching device exceeds the upper limit threshold or is lower than the set difference of the ambient temperature, and use interpolation to complete the missing data. Based on the operating data, the power change rate, junction temperature change rate, cumulative thermal stress index and device temperature fluctuation amplitude are calculated, wherein the cumulative thermal stress index is the temperature-time accumulation of each switching device operating at a junction temperature above a set reference temperature. The input voltage, output current, and instantaneous conversion power are normalized, and the junction temperature of each switching device is standardized to form the multidimensional operating characteristic matrix.

3. The method for optimizing the configuration of power converters in a distribution network based on scenario clustering according to claim 1, characterized in that, Step S2 includes: Based on mutual information coefficients and feature importance, feature filtering is performed on the multidimensional running feature matrix to obtain a dimensionality-reduced feature matrix; K-means clustering is performed on the fast-changing features in the reduced-dimensional feature matrix to obtain the power transient mode, hierarchical clustering is performed on the slow-changing features to obtain the heat load state, and DBSCAN density clustering is performed on the diurnal-changing features to obtain the typical daily mode. These are combined to form a preliminary scene category. Calculate and filter representative indicators for each of the preliminary scenario categories. The representative indicators are obtained by weighted summation of the sample size ratio, thermal damage ratio, and clustering quality. Extreme samples with extreme temperatures, rapid heating, and uneven temperature are identified. After sub-clustering of various extreme samples, they are forcibly included to form the typical transformation scenarios and their characteristic centroid vectors, thermal stress characteristic parameters, and scenario transition probabilities.

4. The method for optimizing the configuration of power converters in a distribution network based on scenario clustering according to claim 1, characterized in that, The step S4, predicting the cumulative thermal stress of the device over a future time period under the current topology, includes: Construct an LSTM-CNN hybrid neural network model, with the input being a multidimensional feature sequence containing a sliding time window of a set past duration; The LSTM-CNN hybrid neural network model includes a two-layer bidirectional LSTM branch, a multi-scale parallel convolutional CNN branch, an attention fusion layer, and a fully connected prediction layer. The forward and backward hidden states of the two-layer bidirectional LSTM branch are concatenated to obtain a temporal feature vector. The multi-scale parallel convolutional CNN branch includes a short-term convolutional path, a medium-term convolutional path, and a long-term convolutional path. The outputs of the three convolutional paths are pooled and concatenated to obtain a spatial feature vector. The attention fusion layer calculates the dynamic weights of the temporal feature vector and the spatial feature vector and performs weighted fusion. The fully connected prediction layer outputs a multi-device junction temperature prediction matrix for multiple future time steps. Based on the multi-device junction temperature prediction matrix, the predicted peak junction temperature and the predicted cumulative thermal damage are calculated. The predicted cumulative thermal damage is the integral of the predicted junction temperature raised to a set power over time when it is above the reference temperature.

5. The method for optimizing the configuration of power converters in a distribution network based on scenario clustering according to claim 4, characterized in that, Step S4, which involves querying the scene-topology adaptation matrix to obtain the accumulated thermal stress of the candidate topologies, includes: The average power, average output current, and average power factor for future time periods are calculated based on the multi-device junction temperature prediction matrix. Calculate the Mahalanobis distance between the prediction parameter vector composed of the average power, the average output current and the average power factor and the feature centroid vector of each typical conversion scenario, and identify the scenario with the smallest Mahalanobis distance as the matching scenario; The maximum junction temperature peak value of each candidate topology corresponding to the matching scenario is queried in the scenario-topology adaptation matrix, and the cumulative thermal stress of each candidate topology is estimated based on the maximum junction temperature peak value. The difference between the predicted cumulative thermal damage of the current topology and the cumulative thermal stress of each candidate topology is calculated as the switching benefit of switching to each candidate topology.

6. The method for optimizing the configuration of power converters in a distribution network based on scenario clustering according to claim 5, characterized in that, The calculation of the switching cost of the topology handover process in step S4 includes: The total switching energy cost is calculated by summing the energy discharged by the bus capacitor during the shutdown phase, the energy charged by the output filter capacitor during the dead zone phase, and the additional energy loss due to efficiency reduction during the soft start phase. The equivalent temperature rise is obtained by dividing the total switching energy cost by the product of the device heat capacity and the number of devices, and is used as the switching cost. A rolling prediction mechanism is adopted, and predictions are made at the current time, the first delay time after the current time, and the second delay time after the current time. When all three predictions determine that the switch to the same alternative topology and the net benefit is greater than the set threshold, the switching process of power reduction, safe disconnection, topology reconstruction, soft start and stability confirmation is executed. During the switching process, the power reduction phase linearly reduces the PWM modulation to the set ratio, the safety disconnect phase blocks the gate drive and waits for the dead time, the topology reconfiguration phase configures the PWM channel mapping and switching frequency of the new topology, and the soft start phase linearly increases the modulation to the target value.

7. A power converter optimization configuration system for power distribution networks based on scenario clustering, characterized in that, The distribution network power converter optimization configuration method based on scenario clustering as described in any one of claims 1-6 is used to implement the distribution network power converter optimization configuration system based on scenario clustering, which includes: The extraction module is used to collect the input voltage, output current, instantaneous conversion power, and junction temperature of each switching device of the power converter, calculate the power change rate and cumulative thermal stress index, and form a multi-dimensional operating characteristic matrix. The clustering module is used to perform multi-level clustering on the multi-dimensional operating feature matrix, and to obtain typical conversion scenarios and their thermal stress characteristic parameters by screening representative indicators of the scenarios. The evaluation module is used to evaluate the thermal performance of multiple alternative topologies under the typical conversion scenarios and construct a scenario-topology adaptation matrix. This includes: designing multiple alternative topologies for the same power conversion function; establishing IGBT switching characteristic models and fourth-order thermal network models for each alternative topology in a simulation platform; substituting the operating parameters corresponding to the feature centroid vectors of each typical conversion scenario into the simulation models of each alternative topology; extracting the maximum junction temperature peak, average junction temperature, inter-device temperature variance, total power loss, and conversion efficiency in the steady-state phase after running for a set duration; constructing a multi-objective comprehensive scoring function based on the maximum junction temperature peak, the conversion efficiency, the inter-device temperature variance, and the output voltage harmonic distortion rate; calculating the comprehensive score of each alternative topology under each typical conversion scenario; and constructing the scenario-topology adaptation matrix with the typical conversion scenarios as rows, the alternative topologies as columns, and the comprehensive score as matrix elements. The query module is used to predict the cumulative thermal stress of devices in the future time period under the current topology, query the scenario-topology adaptation matrix to obtain the cumulative thermal stress of candidate topologies, and use the difference between the two as the switching benefit. The energy loss of device turn-off transient, inrush current and restart process during the topology switching process is converted into equivalent temperature rise as the switching cost. When the net benefit of the switching benefit minus the switching cost is greater than a set threshold in three consecutive rolling prediction cycles, the topology switching is performed.

8. The system according to claim 7, characterized in that, The system collects the input voltage, output current, instantaneous conversion power, and junction temperature of each switching device of the power converter, calculates the power change rate and cumulative thermal stress index, and constructs a multi-dimensional operating characteristic matrix, including: The power converter's operating data is continuously collected at a set sampling interval. The operating data includes the input voltage, the output current, the instantaneous conversion power, the power factor, the junction temperature of each switching device, the heat sink temperature, the ambient temperature, and the cooling fan speed. Remove outliers from the operating data where the junction temperature of each switching device exceeds the upper limit threshold or is lower than the set difference of the ambient temperature, and use interpolation to complete the missing data. Based on the operating data, the power change rate, junction temperature change rate, cumulative thermal stress index and device temperature fluctuation amplitude are calculated, wherein the cumulative thermal stress index is the temperature-time accumulation of each switching device operating at a junction temperature above a set reference temperature. The input voltage, output current, and instantaneous conversion power are normalized, and the junction temperature of each switching device is standardized to form the multidimensional operating characteristic matrix.

Images