A multi-source aware power load distribution platform

Abstract

The present invention discloses a multi-source sensing power load distribution platform, which relates to the field of power distribution technology and includes: a forward normalization processing module for determining power sensing data; a conditional initialization module for determining target uncertainty conditions; a model training module for supervising the training of load distribution models; and a power load distribution management module for managing power load distribution for local power grids. The present invention solves the technical problem of poor load distribution optimization and unbalanced resource utilization caused by the failure to consider uncertainty factors between power equipment and load terminals during power load distribution. The present invention achieves real-time power load distribution optimization through multi-source sensing data, adapting to different load changes and power demands, improving power load distribution efficiency, reducing power line losses, and enhancing the operational stability of local power grids.

Description

Technical Field

[0001] The present invention relates to the technical field of power distribution, and in particular to a multi-source-aware power load distribution platform. Background Art

[0002] With the transformation of the global energy structure and the continuous development of smart grid technology, the power system faces unprecedented challenges. Traditional power distribution processes suffer from problems such as a single data source, delayed information transmission, and uneven load distribution. These problems lead to low power resource utilization efficiency, poor power supply security, and an inability to effectively address the uncertainties brought about by the large-scale integration of renewable energy. In local power grids, dynamic adjustment of power load distribution becomes particularly difficult due to the complex influence of multiple factors such as power equipment production capacity, load characteristics, and local grid topology. Currently, most power load distribution strategies rely on traditional scheduling methods based on fixed parameters and rules. These methods lack effective integration and analysis of multi-source data and fail to fully consider the uncertainties of power equipment and loads. This results in low load scheduling efficiency, poor grid operation stability, and ineffective guarantees for the rational allocation of power resources. Summary of the Invention

[0003] The present application provides a multi-source perception power load distribution platform, which is used to solve the technical problems of poor load distribution optimization effect and unbalanced resource utilization due to the failure to consider the uncertainty factors of power equipment and load ends in the power load distribution process.

[0004] The present application provides a multi-source-aware power load distribution platform, the platform comprising:

[0005] A forward normalization processing module is used to connect a multi-source data interface and determine power perception data, wherein the power perception data includes first power sampling data determined by calling multi-source data related to source and load and performing downsampling fusion, and second power index data determined by performing forward normalization processing on the power index; a condition initialization module is used to introduce source-load uncertainty conditions, traverse the local power grid topology, perform condition initialization based on the production capacity specificity of the power equipment end and the energy consumption specificity based on the load end, and determine the target uncertainty conditions; a model training module is used to combine the local power grid topology with the target uncertainty conditions to supervise the training of the load distribution model, wherein the local power grid topology is marked with power line loss, and load distribution optimization training is performed using lightweight processing and rasterization processing based on expansion and causality; a power load distribution management module is used to transmit the power perception data to the load distribution model, make decisions to determine the power load distribution strategy, and perform power load distribution management on the local power grid.

[0006] In a possible implementation, the model training module also performs the following processing: calling load distribution samples, and supervising the training of the large distribution model with the local power grid topology, the load distribution samples include power sample data-uncertain condition samples-power distribution samples; based on the preset loss degree, the large distribution model is lightweight processed, the expansion grid and the causal grid are divided, and the load distribution model is supervised and trained.

[0007] In a possible implementation, the model training module further performs the following processing: determining a lightweight transfer probability based on the preset loss degree, wherein the lightweight transfer probability is the logical transfer probability of the elements of the large distribution model; dividing the expansion grid based on the preset scale characteristics under the time series and spatial distribution, wherein each grid causality has the same expansion standard; dividing the causal grid based on the power causal relationship, wherein each causal grid has the same power causal relationship; combining the lightweight transfer probability, the expansion grid and the causal grid, performing supervised lightweight supervised training on the large distribution model to determine the load distribution model.

[0008] In a possible implementation, the model training module further performs the following processing: obtaining a first load distribution sample, transmitting it to the distribution large model, determining a first soft sample, where the first soft sample includes a distribution logic chain; traversing the distribution logic chain, performing link node identification and grid labeling based on an expansion grid and a causal grid, and determining a first labeled soft sample; and performing a lightweight training based on the lightweight transfer probability and combining the first load distribution sample and the first soft sample.

[0009] In a possible implementation, the power load distribution management module also performs the following processing: using multi-energy mutual assistance as the first load distribution condition, wherein the multi-energy distribution priority is used as a constraint; using multi-microgrid mutual assistance as the second load distribution condition, wherein the local distribution priority is used as a constraint; based on the first load distribution condition and the second load distribution condition, assisting in power load distribution decision-making.

[0010] In a possible implementation, the forward standardization processing module also performs the following processing: mining the power indicators of power load distribution, wherein the power indicators are identified with weight values; traversing the power indicators, performing indicator type attribution and indicator attribute forward processing, and determining extremely large power indicators, wherein the indicator types include extremely large, extremely small and interval types; matrixing and integrating the extremely large power indicators and performing standardization processing to determine the second power indicator data.

[0011] In a possible implementation, the forward normalization processing module further performs the following processing: traversing the multi-source data, performing a first downsampling process of the data granularity based on the data correlation, and determining the first downsampling data; traversing the first downsampling data, performing a second downsampling process based on the data dimension, and determining the second downsampling data; and performing homologous data fusion on the second downsampling data to determine the first power sampling data.

[0012] In a possible implementation, the power load distribution management module also performs the following processing: tracking the response of the power load distribution strategy to determine the distribution response data; traversing the distribution response data to verify and trace the abnormal distribution, and determining the cause of the abnormal distribution, wherein the deviation based on the distribution volume and the deviation based on the distribution trend are used as verification standards; and based on the abnormal distribution cause, distribution feedback management of the local power grid is performed.

[0013] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0014] The present application connects a multi-source data interface to determine power perception data, wherein the power perception data includes first power sampling data determined by calling multi-source data related to source and load and performing downsampling fusion, and second power index data determined by performing forward normalization processing on the power index; introducing source-load uncertainty conditions, traversing the local power grid topology, performing conditional initialization based on the production capacity specificity of the power equipment end and the energy consumption specificity based on the load end, and determining the target uncertainty conditions; combining the local power grid topology with the target uncertainty conditions, supervising the training of the load distribution model, wherein the local power grid topology is marked with power line loss, and optimizing the load distribution training by lightweight processing and rasterization processing based on expansion and causality; transmitting the power perception data to the load distribution model, making decisions to determine the power load distribution strategy, and performing power load distribution management on the local power grid. The present invention solves the technical problem that in the process of power load distribution, the uncertainty factors between power equipment and load ends are not taken into consideration, resulting in poor load distribution optimization effect and unbalanced resource utilization. Through real-time power load distribution optimization based on multi-source perception data, it can adapt to different load changes and power demands, thereby achieving the technical effects of improving power load distribution efficiency, reducing power line losses and enhancing the operational stability of the local power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0016] Figure 1 A schematic diagram of the structure of a multi-source sensing power load distribution platform provided in an embodiment of the present application;

[0017] Figure 2 A flow chart of a training load distribution model for a multi-source-aware power load distribution platform provided in an embodiment of the present application.

[0018] Explanation of the reference numerals: forward normalization processing module 11 , condition initialization module 12 , model training module 13 , power load distribution management module 14 . DETAILED DESCRIPTION

[0019] This application provides a multi-source perception power load distribution platform to solve the technical problems of poor load distribution optimization and unbalanced resource utilization caused by failure to consider uncertainty factors between power equipment and load ends during the power load distribution process. Through real-time power load distribution optimization based on multi-source perception data, it can adapt to different load changes and power demands, thereby achieving the technical effects of improving power load distribution efficiency, reducing power line losses and enhancing the operational stability of local power grids.

[0020] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only some of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0021] It should be noted that any variations of the terms "include" and "have" are intended to cover non-exclusive inclusions. For example, a process, method, platform, product or server 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 modules that are not explicitly listed or are inherent to these processes, methods, products or devices.

[0022] Examples, such as Figure 1 As shown, an embodiment of the present application provides a multi-source-aware power load distribution platform, which includes:

[0023] The forward normalization processing module 11 is used to connect the multi-source data interface and determine the power sensing data, wherein the power sensing data includes the first power sampling data determined by calling the multi-source data related to the source and load and performing downsampling fusion, and the second power indicator data determined by performing forward normalization processing on the power indicator.

[0024] In this embodiment of the present application, a forward normalization processing module is used to connect to a multi-source data interface to generate power perception data. This power perception data consists of two parts: first power sampling data and second power indicator data. The first power sampling data is generated by invoking multi-source data related to source and load, and undergoing a two-stage downsampling and homogenous data fusion process. The second power indicator data is generated by mining, classifying, forward processing, and matrixing power indicators. Together, these two parts constitute the power perception data.

[0025] Furthermore, in the platform provided in the application embodiment, the forward normalization processing module 11 is further used to:

[0026] The multi-source data is traversed, and based on the data correlation, a first downsampling process of the data granularity is performed to determine the first downsampling data; the first downsampling data is traversed, and a second downsampling process is performed based on the data dimension to determine the second downsampling data; and the second downsampling data is fused with the same source data to determine the first power sampling data.

[0027] In this embodiment, multi-source data is first traversed and a first downsampling process is performed based on correlation. During this stage, multi-source data is collected, including capacity data from power generation equipment (such as real-time generated power), load data from power consumption equipment (such as instantaneous load), and external environmental data (such as temperature, humidity, and wind speed). This data is preprocessed to remove outliers (such as negative generated power) and duplicates. The Pearson correlation coefficient is then used to calculate the correlation between each data point and the target variable (such as load change or power line loss). Data with high correlations typically exhibit strong statistical connections. Pearson correlation coefficients close to 1 (positive correlation) or -1 (negative correlation) are considered highly correlated. For example, a correlation coefficient of 0.85 between generated power and load change is considered highly correlated. Data close to 0, such as a correlation coefficient of 0.1 between ambient humidity and line loss, are considered weakly correlated and are eliminated. During this process, data granularity is based on the original acquisition frequency (seconds, minutes, and hours). By screening key data sets with a correlation threshold greater than 0.5, the first downsampling data is generated, covering key operational characteristics of the power system.

[0028] The first downsampled data is then traversed and a second downsample is performed based on principal component analysis (PCA). The first downsampled data is organized into a feature matrix, where each column represents a feature dimension, such as power generation, load factor, and line loss rate. The PCA algorithm is used to calculate the variance contribution of each feature dimension. For example, the contribution of power generation is 40%, the contribution of load factor is 30%, the contribution of line loss rate is 20%, and the contribution of ambient temperature is only 5%. A threshold is set based on the cumulative variance contribution (e.g., reaching 90%), and the feature dimension with the highest contribution is selected. Dimensions with lower overall information contribution (such as ambient temperature) are eliminated. This process extracts the key features that have the greatest impact on power load distribution, reduces dimensional redundancy, and generates more refined second downsampled data.

[0029] Finally, the second downsampled data undergoes homologous data fusion. Similar data from different sources within the dataset (e.g., voltage values measured by multiple sensors) is fused using a weighted average method. Weights are assigned based on the reliability of the data source; for example, a high-precision voltage sensor is weighted 0.7, while a standard sensor is weighted 0.3. During the fusion process, a weighted average is calculated for each set of similar data to eliminate bias caused by sensor differences or measurement errors, ensuring data consistency and accuracy. Ultimately, highly consistent first-level power sampling data is generated.

[0030] Furthermore, in the platform provided in the application embodiment, the forward normalization processing module 11 is further used to:

[0031] Mining power indicators of power load distribution, wherein the power indicators are marked with weight values; traversing the power indicators, performing indicator type attribution and indicator attribute forward processing, and determining extremely large power indicators, wherein the indicator types include extremely large, extremely small, and interval types; matrix-integrating the extremely large power indicators and performing standardization processing to determine the second power indicator data.

[0032] In an embodiment of the present application, the power indicators of power load distribution are first mined. Core features such as voltage, current, power factor, line loss rate and load rate are collected from the power grid operation data. The relationship between each feature and the load distribution efficiency is calculated using a correlation analysis method (such as the Pearson correlation coefficient). For example, the correlation coefficient of the power factor is 0.8, the correlation coefficient of the load rate is 0.7, and the correlation coefficient of the ambient temperature is only 0.2. The threshold is set to 0.5, and the features with a correlation higher than this threshold are used as core indicators of power load distribution. Then, a weight value is assigned to each core indicator to reflect its importance in load distribution, for example, the weight of the power factor is 0.4, and the weight of the line loss rate is 0.3.

[0033] Next, the mined power indicators are traversed and assigned to different types. Each indicator is classified using threshold rules. Power factor and power generation capacity are classified as extreme indicators (larger values are preferred), line loss rate and equipment failure rate are classified as extreme indicators (smaller values are preferred), and load factor is classified as an interval indicator (optimal values are between 70% and 90%). Subsequently, the indicator attributes are forward-processed. Different forward-processing methods are used for different types of indicators. For extreme indicators, such as line loss rate, a reciprocal method is used to convert a line loss rate of 0.02 to a positive value of 50. For interval indicators, such as load factor, an interval mapping method is used to map values within the optimal range (70%-90%) to high scores. For example, 80% is mapped to 0.9, and 50% is mapped to 0.3. For extreme indicators, their values remain unchanged, as they already have positive characteristics. Through this process, all indicators are converted into extreme power indicators with a unified optimization direction, ensuring a consistent analysis benchmark for different types of indicators.

[0034] The large-scale power indicators are then matrix-integrated and standardized. All forward-processed large-scale power indicators are combined into a matrix, where each column represents an indicator and each row represents a time point or device status. The maximum and minimum value normalization method is then applied to this matrix to normalize all indicator values to the same numerical range, ensuring that different indicators have a unified dimension.

[0035] Through the above steps, the whole process from mining and forward processing of power indicators to matrix integration and standardized processing is completed to generate the second power indicator data.

[0036] The condition initialization module 12 is used to introduce source-load uncertainty conditions, traverse the local power grid topology, perform condition initialization based on the power equipment end's production capacity specificity and the load end's energy consumption specificity, and determine the target uncertainty conditions.

[0037] In an embodiment of the present application, first, the conditional initialization module introduces the source-load uncertainty condition through real-time data extraction and simple statistical analysis. The capacity specificity of the source end is directly obtained from the real-time operating data of the power generation equipment, including the current power generation power and its historical maximum and minimum values. The range of power generation fluctuations is calculated and defined as the capacity uncertainty condition of the source end. At the same time, the energy demand of the load end is extracted from the real-time power consumption record of the load equipment. By calculating the mean and maximum value of the power consumption data, the load fluctuation range is obtained to form the energy uncertainty condition of the load end. Finally, the fluctuation conditions of the source end and the load end are integrated to form a preliminary source-load uncertainty condition.

[0038] Next, the topology of the local power grid is traversed using graph theory. The nodes (e.g., power stations, load centers) and lines (e.g., cables, transformers) of the local power grid are represented as a graph model, with nodes as vertices and lines as edges. Each node is traversed one by one, extracting the node's associated power generation, load, and line transmission capacity data. The difference between the node's power generation capacity and the load demand is calculated. The transmission capacity and actual load data of each line are also recorded, identifying key nodes and lines that may be affected by source-load fluctuations.

[0039] After completing the topology analysis, conditional initialization is performed directly based on the specificity of energy production and energy consumption. This initialization is achieved by mapping the fluctuation range of power generation equipment to power generation nodes and the fluctuation range of load equipment demand to load nodes. Initialization conditions are generated by directly linking the uncertainty conditions at the source and load ends, combined with the transmission constraints between nodes in the local power grid topology.

[0040] Finally, we screen out key target uncertainty conditions. By examining the initialization results, we directly mark nodes with large power generation fluctuations or drastic changes in load demand, as well as lines with transmission capacity close to the upper limit. These characteristics are then used as target uncertainty conditions.

[0041] The model training module 13 is used to supervise the training of the load distribution model by combining the local power grid topology and the target uncertainty condition, wherein the local power grid topology is marked with power line loss, and the load distribution optimization training is performed by lightweight processing and rasterization processing based on expansion and causality.

[0042] In an embodiment of the present application, the model training module supervises the training of the load distribution model by combining the local power grid topology with the target uncertainty conditions. The local power grid topology identifies the power line loss characteristics as the core constraint conditions for model training. The power line loss is preset, for example, a loss of 0.003% per meter, that is, during the power grid transmission process, each meter of transmission distance will result in a power loss of 0.003%. This characteristic is embedded in the model as a constraint parameter to correct the impact of transmission loss in the distribution strategy. Combined with the load distribution sample data (including power sample data, uncertainty condition samples and power distribution samples), a large distribution model is constructed and optimized. Lightweight processing technology is used to determine the lightweight transfer probability to reduce the complexity of the model; at the same time, based on the expansion grid and causal grid division technology, the local characteristics and causal relationships of the power grid operation are captured to optimize the load distribution strategy of the model. Finally, an optimized load distribution model is output.

[0043] Further, such as Figure 2 As shown, in the platform provided by the application embodiment, the model training module 13 is also used to:

[0044] Calling load distribution samples, and using the local power grid topology to supervise the training of the distribution model, the load distribution samples include power sample data - uncertain condition samples - power distribution samples; based on the preset loss degree, the distribution model is lightweight processed, divided into expansion grid and causal grid, and supervised training of the load distribution model.

[0045] In an embodiment of the present application, first, the load distribution sample is called and combined with the local power grid topology to build a large distribution model. The load distribution sample includes power sample data (such as voltage, current, power factor, etc.), uncertain condition samples (such as the source end power generation fluctuation range and the load end demand change range) and power distribution samples (such as load distribution results under historical conditions). The local power grid topology provides core information such as node distribution, power transmission path and power line loss, which are used as constraints for model training. For example, the topology clarifies the connection relationship and path loss rate (such as 0.003% per meter) between each node in the power grid. After these data are organized as input features and training labels, a supervised learning framework (such as random forest) is used to build a large distribution model.

[0046] The large allocation model is then lightweighted based on a preset loss. By analyzing the logical transfer paths within the model, the contribution of each path to the output is calculated, and feature importance analysis methods (such as SHAP values) are used to quantify the path contribution. A pruning threshold is set based on the preset loss, retaining only high-contribution paths and removing low-contribution paths, thereby reducing model complexity while maintaining the model's ability to express core features.

[0047] Subsequently, the expansion grid and causal grid are divided to optimize model training. For the expansion grid, nodes with similar operating characteristics are clustered using clustering algorithms (such as K-means) based on time series and spatial distribution characteristics, and these nodes are divided into expansion grids. During training, nodes within the expansion grid share the same characteristic weights, facilitating the optimization of local characteristics. For the causal grid, a causal relationship network is constructed to analyze the causal relationships of power flow. Causal analysis methods (such as Granger causality analysis) are used to assign nodes with direct interactions to the causal grid. The causal grid emphasizes the causal relationships between nodes and can enhance the model's ability to express power transmission logic.

[0048] Finally, supervised training is performed on the large load distribution model. This training combines the lightweight transfer probability, the dilation grid, and the causal grid division results with the load distribution sample data. By continuously optimizing the loss function (such as mean squared error) and adjusting the model parameters, the model gradually learns the mapping relationship between the load distribution sample data and the actual distribution results. This process completes the training of the load distribution model.

[0049] Furthermore, in the platform provided in the application embodiment, the model training module 13 is also used to:

[0050] Based on the preset loss degree, a lightweight transfer probability is determined, wherein the lightweight transfer probability is the logical transfer probability of the elements of the large distribution model; based on the preset scale characteristics under the time series and spatial distribution, the expansion grid is divided, wherein each grid causality has the same expansion standard; based on the power causal relationship, the causal grid is divided, wherein each causal grid has the same power causal relationship; combining the lightweight transfer probability, the expansion grid and the causal grid, the large distribution model is supervised lightweight supervised training to determine the load distribution model.

[0051] In an embodiment of the present application, the lightweight transfer probability is first determined based on a preset loss degree. The preset loss degree defines the range of performance accuracy loss allowed during the model simplification process, and guides the model lightweight processing. The contribution of each logical path within the large model to the output result is calculated and distributed through a feature importance analysis method (such as the SHAP value) to obtain a contribution score for the path. Then, the logical paths are sorted from high to low according to the contribution, and a pruning threshold is set according to the preset loss degree. Paths with contributions above the threshold are retained, and paths below the threshold are sparsely processed. The lightweight transfer probability is calculated by dividing the contribution by the maximum contribution, which represents the retention probability of each path, thereby streamlining the logical paths and reducing the complexity of the model while maintaining the ability to express key features.

[0052] The expanded grid is then divided based on preset scale characteristics of time series and spatial distribution. By analyzing the time series operating characteristics (such as load fluctuations and power generation changes) and spatial distribution characteristics (such as geographic location and connectivity) of power grid nodes, scale characteristics such as time window length (such as 1 hour) and spatial coverage are set. Nodes with similar operating characteristics are clustered using clustering algorithms (such as K-means) and assigned to the same expanded grid. Within the same expanded grid, nodes share unified feature weights and optimization objectives, reducing redundant computations between nodes within the grid, capturing the local correlation characteristics of the power grid, and improving model training efficiency.

[0053] The causal grid is then divided based on power causal relationships. By constructing a causal network, the power transmission paths and causal relationships between nodes in the power grid are analyzed. Causal analysis methods (such as Granger causality analysis) are used to determine the direct causal relationships between nodes. For example, whether fluctuations at a power generation node significantly affect changes at downstream load nodes. Nodes with causal relationships are grouped into the same causal grid, forming a causal chain within the grid. Within the causal grid, nodes share the same causal characteristics (such as power flow direction and optimization objectives), strengthening the model's ability to model power transmission logic and thus more accurately express causal paths.

[0054] Finally, the large-scale distribution model is trained using supervised lightweight methods, combining lightweight transfer probabilities, an expansion grid, and a causal grid. Load distribution samples (including power sample data, uncertain condition samples, and distribution samples) are fed into the model along with the aforementioned lightweight transfer probabilities, expansion grid, and causal grid characteristics. The model parameters are optimized using a supervised learning algorithm. In each round of training, the lightweight transfer probabilities guide the model to simplify the path structure, the expansion grid optimizes the expression of local characteristics, and the causal grid enhances the modeling of power transmission logic. By iteratively optimizing the loss function (such as mean squared error), the model gradually converges, ultimately generating an accurate and efficient load distribution model.

[0055] Furthermore, in the platform provided in the application embodiment, the model training module 13 is also used to:

[0056] A first load distribution sample is obtained and transmitted to the distribution large model to determine a first soft sample, where the first soft sample includes a distribution logic chain; the distribution logic chain is traversed to perform link node identification and grid labeling based on an expansion grid and a causal grid to determine a first labeled soft sample; and a lightweight training is performed based on the lightweight transfer probability and the first load distribution sample and the first soft sample.

[0057] In an embodiment of the present application, in order to optimize the distribution of the large model and generate the final load distribution model, samples are obtained one by one from the load distribution sample set, and the model performance is gradually optimized by generating the first soft sample and the first labeled soft sample in combination with lightweight training, and finally an accurate and efficient load distribution model is generated.

[0058] Specifically, the first load distribution sample is obtained. This sample contains power sample data (basic data describing grid operating characteristics, such as node voltage, current, and power factor), uncertain condition samples (such as the fluctuation range of the generation end and the range of load demand changes), and power distribution samples (such as load distribution results under historical conditions). This data is input into the large distribution model, and the first soft sample is generated through the model's initial mapping. The first soft sample contains the distribution logic chain, which describes the logical path of the input features in the model and their transmission relationship, that is, how the input features are transferred layer by layer through the model to produce the target distribution result.

[0059] Next, the distributed logic chain in the first soft sample is traversed and the link nodes are grid-labeled to generate the first labeled soft sample. For the expanded grid labeling, the temporal and spatial characteristics of the logic chain nodes are extracted. Clustering algorithms (such as K-means) are used to group nodes with similar operating characteristics into the same expanded grid and assign them uniform characteristic weights. This process optimizes the local characteristic representation of nodes within the grid and reduces redundant computation. For the causal grid labeling, a causal relationship network is constructed. Causal analysis methods (such as Granger causality analysis) are used to identify the causal paths of nodes in the logic chain. Nodes with direct causal relationships are grouped into the same causal grid and assigned uniform causal path characteristics. This labeling generates the first labeled soft sample, which integrates the information of the logic chain and the grid characteristics.

[0060] The large load distribution model is then trained with lightweight transfer probabilities. The lightweight transfer probabilities are calculated using path contributions, representing the importance of each path in the logic chain to the model output. The probability of path retention is determined by subtracting the maximum contribution from the contribution. High-contribution paths are prioritized, while low-contribution paths are thinned out or removed. The first load distribution sample and its corresponding first labeled soft sample are input into the model, and the model parameters are optimized using a supervised learning algorithm. During training, the expansion grid is used to optimize local feature representation, the causal grid enhances causal path modeling, and lightweight processing simplifies the model structure.

[0061] The above steps are then repeated for each sample in the load distribution sample set, sequentially generating and labeling soft samples from the sample set, and performing lightweight training. Through multiple rounds of training, the large distribution model gradually learns the mapping relationship between input features and target outputs, significantly optimizing the model's logical path and performance.

[0062] Finally, an accurate and efficient load distribution model is generated by training all load distribution samples.

[0063] The power load distribution management module 14 is used to transmit the power sensing data to the load distribution model, make decisions to determine the power load distribution strategy, and perform power load distribution management on the local power grid.

[0064] In an embodiment of the present application, first, the power load distribution strategy is determined based on the load distribution model. The power load distribution management module receives the acquired power sensing data (including real-time power status data and standardized power index data) and transmits it to the load distribution model. The load distribution model parses and calculates the input data through its internal logical path and optimization parameters. Specifically, the load distribution model analyzes the load demand, power generation end fluctuations, uncertain conditions and other characteristics in the power grid, and combines the power grid topology, power line loss and node causality to generate a distribution plan for each power grid node. The determined power load distribution strategy includes a control plan for the output power of the power generation node and a plan to meet the power demand of the load node, aiming to optimize the operating efficiency and stability of the power grid.

[0065] Subsequently, the module manages power load distribution within the local power grid. Based on the generated power load distribution strategy, the module issues specific load dispatch instructions to the power generation and load devices within the local power grid. For example, on the power generation side, these instructions might include adjusting the power output of the generator sets. On the load side, these instructions might prioritize the power needs of critical loads (such as hospitals or industrial equipment) while reducing or delaying the power supply to non-critical loads (such as commercial billboard lighting). This optimizes the overall operational efficiency of the power grid within limited resources.

[0066] Furthermore, in the platform provided in the embodiment of the application, the power load distribution management module 14 is further configured to:

[0067] The first load distribution condition is multi-energy mutual assistance, wherein the multi-energy distribution priority is used as a constraint; the second load distribution condition is multi-microgrid mutual assistance, wherein the local distribution priority is used as a constraint; based on the first load distribution condition and the second load distribution condition, auxiliary power load distribution decision-making is performed.

[0068] In the embodiment of the present application, the first load distribution condition (multi-energy mutual assistance) is first determined. The supply and demand characteristics of various energy forms (such as electricity, wind energy, natural gas, etc.) are extracted, including power generation capacity, energy storage status, and load demand. For energy conversion efficiency, predetermined efficiency parameters are used. These parameters are set based on historical operating data or system calibration values, such as the conversion efficiency of wind power generation and the thermal efficiency of natural gas power generation. By analyzing the energy conversion efficiency and the current supply and demand status, high-efficiency energy forms are prioritized. For example, when electricity demand increases but wind energy resources are sufficient, wind power generation is prioritized to meet demand; when wind power generation is insufficient, natural gas power generation is called upon to supplement it. By integrating the synergistic effects of various energy forms, a multi-energy mutual assistance condition is generated as one of the constraints for load distribution.

[0069] Next, the second load distribution condition (multi-microgrid mutual assistance) is determined. The operating characteristics of the microgrids within the local power grid are analyzed, including their generation capacity (such as photovoltaic and wind power), energy storage status, and load distribution. By prioritizing the dispatch of local resources, the load demands within the microgrid are prioritized. For example, when a microgrid has sufficient generation capacity, its load demand is met primarily by local resources. When a microgrid faces power shortages, mutual assistance is achieved by analyzing the resource reserves of neighboring microgrids, and coordination among multiple microgrids is coordinated to balance power supply and demand. The resulting multi-microgrid mutual assistance condition, based on prioritizing local power distribution, effectively enhances the self-sufficiency and flexibility of the local power grid.

[0070] After determining the first and second load distribution conditions, the power-sensing data is transmitted to the load distribution model using them as constraints. Based on the input power-sensing data and constraints, the load distribution model comprehensively analyzes the grid's operating status, energy characteristics, and load demand to generate an optimized load distribution strategy.

[0071] Finally, the local power grid is managed for load distribution. Specific instructions are issued to power generation and load devices in the grid based on the load distribution strategy, such as requiring wind turbines to increase power output or starting backup natural gas power generation equipment in low wind conditions.

[0072] Furthermore, in the platform provided in the embodiment of the application, the power load distribution management module 14 is further configured to:

[0073] The power load distribution strategy is tracked in response to determine the power distribution response data; the power distribution response data is traversed to verify and trace the abnormal power distribution and determine the abnormal power distribution cause, wherein the deviation based on the power distribution amount and the deviation based on the power distribution trend are used as verification standards; based on the abnormal power distribution cause, the power distribution feedback management of the local power grid is carried out.

[0074] In this embodiment, the power load distribution strategy is first tracked to determine distribution response data. Sensors and smart meters installed at each grid node collect real-time data on actual power distribution, load response, voltage and current fluctuations, and power output of power generation equipment. The collected actual operating data is compared node by node with the target data in the distribution strategy, and the difference is calculated to generate distribution response data.

[0075] Then, the power distribution response data is traversed to verify and trace the abnormal power distribution and determine the cause of the abnormal power distribution. Specifically, first, based on the deviation check of the power distribution, by comparing the actual power distribution with the target power distribution, it is determined whether the difference exceeds the preset deviation range, such as ±10%. If the difference of a certain node exceeds this range, it is determined to be an abnormal node. In addition, based on the deviation check of the power distribution trend, the time series of the power distribution response data is analyzed to determine whether the actual power distribution trend is consistent with the load demand change trend. For example, when the load demand increases, the actual power distribution fails to increase synchronously, and it is marked as a trend abnormality.

[0076] After identifying the abnormal node, we conduct a source tracing analysis, tracing back the power distribution path and examining the operating status and data of related upstream and downstream nodes. For example, if a node experiences insufficient power supply, we can gradually trace back to its power source and transmission path to check for issues such as insufficient power generation equipment output, excessive power line losses, or overloaded loads, thereby identifying the specific cause of the abnormal power distribution.

[0077] Finally, based on the causes of abnormal power distribution, feedback management is conducted on the local power grid's distribution. For short-term anomalies, such as sudden load fluctuations, the distribution strategy is adjusted immediately to quickly replenish power by dispatching backup power generation equipment or adjusting distribution priorities. For long-term anomalies, such as excessive line losses or aging equipment, feedback recommendations are generated, such as optimizing distribution routes, replacing equipment, or upgrading grid infrastructure. These improvements are then incorporated into subsequent distribution strategy adjustments. Through closed-loop feedback management, distribution plans are continuously optimized and grid operational efficiency is improved.

[0078] In the embodiments of the present application, in summary, the embodiments of the present application have at least the following technical effects:

[0079] The present application connects a multi-source data interface to determine power perception data, wherein the power perception data includes first power sampling data determined by calling multi-source data related to source and load and performing downsampling fusion, and second power index data determined by performing forward normalization processing on the power index; introducing source-load uncertainty conditions, traversing the local power grid topology, performing conditional initialization based on the production capacity specificity of the power equipment end and the energy consumption specificity based on the load end, and determining the target uncertainty conditions; combining the local power grid topology with the target uncertainty conditions, supervising the training of the load distribution model, wherein the local power grid topology is marked with power line loss, and optimizing the load distribution training by lightweight processing and rasterization processing based on expansion and causality; transmitting the power perception data to the load distribution model, making decisions to determine the power load distribution strategy, and performing power load distribution management on the local power grid. The present invention solves the technical problem that in the process of power load distribution, the uncertainty factors between power equipment and load ends are not taken into consideration, resulting in poor load distribution optimization effect and unbalanced resource utilization. Through real-time power load distribution optimization based on multi-source perception data, it can adapt to different load changes and power demands, thereby achieving the technical effects of improving power load distribution efficiency, reducing power line losses and enhancing the operational stability of the local power grid.

[0080] It should be noted that the order in which the embodiments of the present application are presented is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. Furthermore, the foregoing descriptions of specific embodiments of this specification are provided. The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential sequence shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0081] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

[0082] This specification and drawings are merely illustrative of the present application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Obviously, those skilled in the art may make various modifications and variations to this application without departing from the scope of this application. Thus, this application is intended to include such modifications and variations as fall within the scope of this application and its equivalents.

Claims

1. A multi-source sensing power load distribution platform, characterized by: The platform includes: a forward normalization processing module, configured to connect to a multi-source data interface and determine power sensing data, wherein the power sensing data includes first power sampling data determined by downsampling and fusing multi-source data related to source and load, and second power index data determined by forward normalization processing of the power index; The conditional initialization module is used to introduce source-load uncertainty conditions, traverse the local power grid topology, perform conditional initialization based on the power equipment-side capacity specificity and the load-side energy specificity, and determine the target uncertainty conditions; a model training module for supervising the training of a load distribution model based on a local power grid topology and the target uncertainty condition, wherein the local power grid topology is identified by power line losses and load distribution optimization training is performed using lightweight processing and rasterization processing based on expansion and causality; An electric load distribution management module, configured to transmit the electric power sensing data to the load distribution model, determine the electric load distribution strategy, and manage the electric load distribution of the local power grid; Wherein, the power load distribution management module is used to: The first load distribution condition is to use multi-energy mutual assistance, where the multi-energy distribution priority is used as a constraint; The second load distribution condition is to use the mutual assistance of multiple microgrids, where the local power distribution priority is used as a constraint; Assisting in making a power load distribution decision based on the first load distribution condition and the second load distribution condition; Wherein, the model training module is used to: Calling load distribution samples to supervise and train a large distribution model based on the local power grid topology, wherein the load distribution samples include power sample data, uncertain condition samples, and power distribution samples; Based on a preset loss degree, the large distribution model is lightweighted, the expansion grid and the causal grid are divided, and the load distribution model is supervised and trained; Wherein, the model training module is used to: Based on the preset loss degree, determining a lightweight transfer probability, wherein the lightweight transfer probability is a logical transfer probability of elements of the large allocation model; Based on the preset scale characteristics of time series and spatial distribution, the expansion grid is divided, where each grid has the same expansion standard within the cause and effect; Based on the power causal relationship, the causal grid is divided, wherein each causal grid has the same power causal relationship; Combine the lightweight transfer probability, the expansion grid and the causal grid to perform supervised lightweight supervised training on the large distribution model to determine the load distribution model; Wherein, the model training module is used to: Obtaining a first load distribution sample, transmitting the sample to the distribution large model, and determining a first soft sample, wherein the first soft sample includes a distribution logic chain; Traversing the distribution logic chain, performing link node identification and grid marking based on the expansion grid and the causal grid, and determining a first marked soft sample; Taking the lightweight transfer probability as a benchmark, a lightweight training is performed in combination with the first load distribution sample and the first soft sample.

2. The multi-source sensing power load distribution platform according to claim 1, characterized in that: The forward normalization processing module is used to: mining power indicators for power load distribution, wherein the power indicators are identified with weight values; Traversing the power indicators, performing indicator type classification and indicator attribute forward processing, and determining the extremely large power indicators, wherein the indicator types include extremely large, extremely small, and interval types; The extremely large power indicators are integrated in a matrix manner and standardized to determine the second power indicator data.

3. The multi-source sensing power load distribution platform according to claim 1, characterized in that: The forward normalization processing module is used to: Traversing the multi-source data, performing a first downsampling process of the data granularity based on the data correlation, and determining first downsampling data; Traversing the first down-sampled data, performing a second down-sampling process based on the data dimension, and determining second down-sampled data; The second down-sampled data is subjected to homologous data fusion to determine the first power sampling data.

4. The multi-source sensing power load distribution platform according to claim 1, characterized in that: The power load distribution management module is further used to: Response tracking of the power load distribution strategy is performed to determine power distribution response data; Traversing the power distribution response data, verifying and tracing the abnormal power distribution, and determining the cause of the abnormal power distribution, wherein the deviation based on the power distribution amount and the deviation based on the power distribution trend are used as verification criteria; Based on the abnormal power distribution inducement, power distribution feedback management of the local power grid is performed.

Images