A power distribution network load forecasting and dispatching system based on multi-modal data and semi-modal learning

Abstract

This invention relates to the field of intelligent operation technology for distribution networks, specifically disclosing a distribution network load forecasting and dispatching system based on multimodal data and semimodal learning. The system collects multimodal data from the distribution network in real time, classifying monitoring nodes into labeled and unlabeled nodes. It utilizes labeled node data to construct a distribution network causal knowledge base containing causal relationships between load type and meteorological conditions. The causal relationships are used as prior knowledge to configure a semi-supervised meta-learning model including a main forecasting network and a knowledge query network. Collaborative training is performed using labeled data as the supervisory signal and unlabeled data as a supplement. During the forecasting phase, the initial load forecast from the main forecasting network and the target causal relationships retrieved by the knowledge query network are combined to calculate and implement physical law compliance corrections. Based on the corrected load forecast results, multi-level safety checks are performed, and a graded dispatching strategy is automatically executed according to the warning level.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation technology of distribution networks, specifically to a distribution network load forecasting and scheduling system based on multimodal data and semimodal learning. Background Technology

[0002] With the continuous expansion of power distribution networks and the increasing penetration of renewable energy, the volatility and uncertainty of power load are becoming increasingly pronounced. In recent years, multimodal data fusion technology has provided new insights for load forecasting. By combining electricity load data, meteorological environmental data, and power grid topology data, a more comprehensive load profile can be constructed. However, existing methods often face problems such as uneven data quality and scarce labels when fusing multimodal data. Many monitoring nodes in the distribution network lack accurate historical load labels due to outdated equipment or incomplete monitoring, making it difficult to apply supervised learning methods. Furthermore, most existing forecasting models remain at the data level of correlation analysis, failing to integrate the physical laws of the power system into the forecasting process in the form of structured knowledge. This often leads to biases in forecasts that contradict common sense under extreme weather conditions.

[0003] The existing technology has the following shortcomings:

[0004] Existing load forecasting methods for distribution networks suffer from fundamental limitations in their structure and learning paradigms: purely data-driven models based on statistical correlations cannot incorporate the physical laws of the power system, leading to prediction biases that violate physical principles under data scarcity or extreme operating conditions; simultaneously, traditional supervised learning paradigms struggle to overcome the limitations of labeled data and cannot effectively utilize the large amount of unlabeled node data existing in the distribution network, resulting in severely insufficient generalization ability of the models under complex climatic conditions. These two interconnected defects together contribute to the vulnerability of forecasting systems in real-world applications, affecting not only forecast accuracy but also posing potential risks to grid security dispatching. Summary of the Invention

[0005] The purpose of this invention is to provide a power distribution network load forecasting and scheduling system based on multimodal data and semimodal learning to solve the problems mentioned above.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A distribution network load forecasting and dispatching system based on multimodal data and semimodal learning includes:

[0008] The multimodal data acquisition and processing module is used to acquire multimodal data from multiple monitoring nodes in the power distribution network in real time and divide the monitoring nodes into labeled nodes and unlabeled nodes.

[0009] The causal knowledge base construction module establishes a causal knowledge base for the distribution network based on multimodal data with labeled nodes, which includes the causal relationship between load type and meteorological conditions.

[0010] The semi-supervised meta-learning model configuration module uses causal relationships as prior knowledge to configure a semi-supervised meta-learning model that includes a main prediction network and a knowledge query network.

[0011] The collaborative training module uses multimodal data with labeled nodes as supervision signals and multimodal data without labeled nodes as supplementary data to train the knowledge query network to retrieve relevant causal relationships from the causal knowledge base.

[0012] The load forecasting and correction module generates an initial load forecast through the main forecasting network, while using a knowledge query network to retrieve the target causal relationships under the current meteorological conditions, and performs physical law conformity correction on the initial load forecast.

[0013] The scheduling strategy generation and execution module formulates and executes the distribution network scheduling strategy based on the final load forecast results.

[0014] As a further aspect of the present invention: the division of monitoring nodes into tagged nodes and untagged nodes specifically includes:

[0015] Based on the historical data integrity rate, data anomaly fluctuation frequency, and equipment operation stability indicators of each monitoring node, a real-time credit score is calculated for each monitoring node.

[0016] Based on the real-time credit score, the monitoring nodes are classified and processed. Monitoring nodes with a real-time credit score greater than or equal to the first threshold are automatically classified as labeled nodes, and monitoring nodes with a real-time credit score less than the first threshold are classified as unlabeled nodes.

[0017] As a further aspect of the present invention: the establishment of a distribution network causal knowledge base containing the causal relationship between load type and meteorological conditions specifically includes:

[0018] Based on historical multimodal data with labeled nodes, candidate causal relationships between load types and meteorological conditions are identified, with potential causal relationships discovered by analyzing the temporal relationship and statistical dependence between load changes and meteorological parameter changes.

[0019] The candidate causal associations were verified in multiple dimensions, including verifying the temporal consistency, association strength and physical rationality between load changes and meteorological parameter changes, and screening out the reliable causal associations that passed the verification.

[0020] The verified reliable causal relationships are classified and stored according to load type and meteorological conditions to establish a structured causal knowledge base. The confidence score of each causal relationship is calculated to form a complete causal knowledge base for the power distribution network.

[0021] As a further aspect of the present invention: the use of causal relationships as prior knowledge, and the configuration of a semi-supervised meta-learning model comprising a main prediction network and a knowledge query network, specifically includes:

[0022] Based on the causal relationships in the causal knowledge base, a query template library is constructed, where each query template corresponds to a mapping relationship from a type of meteorological conditions to a type of load and contains a complete semantic description framework;

[0023] Based on the structural characteristics of the query template library, a parameter configuration table for the knowledge query network is generated. The parameter configuration table defines the connection weight initialization scheme and feature extraction range for each layer of the knowledge query network.

[0024] The component structures of the knowledge query network and the main forecast network are initialized based on the parameter configuration table. The knowledge query network is responsible for receiving meteorological features and outputting knowledge base query instructions, while the main forecast network is responsible for processing multimodal data and generating load forecasts.

[0025] As a further aspect of the present invention: the trained knowledge query network retrieves relevant causal relationships from the causal knowledge base, specifically including:

[0026] A supervised training sample set is constructed based on the meteorological feature sequences of labeled nodes and their corresponding load type labels;

[0027] The supervised training sample set is input into the knowledge query network for forward computation to obtain the query command output by the network, and the query command is matched with the actual causal relationship in the causal knowledge base for similarity.

[0028] The loss function is calculated based on the similarity matching results, and the parameters of the knowledge query network are adjusted through backpropagation so that the query instructions output by the knowledge query network can accurately point to the load type associated with the input meteorological characteristics.

[0029] As a further aspect of the present invention: the step of calculating the loss function based on the similarity matching result and adjusting the parameters of the knowledge query network through backpropagation specifically includes:

[0030] Based on the matching scores between query commands and various causal relationships in the causal knowledge base, a multi-level similarity distribution map is constructed, where the first level measures the degree of matching of semantic features and the second level evaluates the degree of matching of structural features.

[0031] The differential loss function is calculated based on the multi-level similarity distribution map. Positive reinforcement is applied to the matching score of correct causal associations, negative suppression is applied to the matching score of incorrect associations, and the discriminativeness of the loss function is enhanced by the differential amplification factor.

[0032] Gradient information is calculated based on the differential loss function, and the connection weights and feature extraction parameters of the knowledge query network are adjusted through the backpropagation process.

[0033] As a further aspect of the present invention: the physical law conformity correction of the initial load forecast specifically includes:

[0034] Based on the target causal relationships obtained from knowledge query network retrieval, a dynamic influence relationship diagram between load type and meteorological parameters is constructed.

[0035] Based on the influence intensity of each load type in the dynamic influence relationship diagram, calculate the correction factors for each type of load under the current meteorological conditions;

[0036] The correction factor is weighted and fused with the initial load forecast generated by the main forecast network to generate a final load forecast that conforms to physical laws.

[0037] As a further aspect of the present invention: the process of obtaining the correction factor is as follows:

[0038] Based on the influence intensity of each load type in the dynamic influence relationship diagram, the degree of deviation of the current meteorological parameters from the historical baseline value is obtained;

[0039] The degree of deviation and the influence intensity of the corresponding load type are normalized to obtain the sensitivity coefficient of each load type under the current meteorological conditions;

[0040] Based on the sensitivity coefficient and the preset correction benchmark value, the correction factor for various types of loads is calculated through weighted product.

[0041] As a further aspect of the present invention: the generation of the final load prediction that conforms to physical laws specifically includes:

[0042] A dynamic weight allocation table is established based on the correction factors for various types of loads, and the weight ratio of various types of loads in the final forecast is determined according to the magnitude of the correction factors.

[0043] Based on the weight ratios in the dynamic weight allocation table, the initial load forecast is decomposed into forecast components of different load types, and corresponding weight adjustments are made to each forecast component.

[0044] The weighted load type forecast components are recombined to generate a final load forecast that conforms to physical laws.

[0045] As a further aspect of the present invention: the step of formulating and executing a distribution network dispatching strategy based on the final load forecast results specifically includes:

[0046] Multi-level safety verification of power grid operation status is carried out based on the final load forecast results, including line load rate assessment, transformer capacity margin analysis and node voltage stability calculation.

[0047] Based on the safety verification results, a graded early warning signal is generated. When the predicted load exceeds the first threshold of the equipment's rated capacity, preventive scheduling is initiated. When it exceeds the second threshold, emergency scheduling is initiated.

[0048] Perform dispatching operations corresponding to the warning level. Preventive dispatching includes adjusting transformer taps and switching reactive power compensation devices, while emergency dispatching includes activating load shedding plans and adjusting the output of distributed power sources.

[0049] The beneficial effects of this invention are:

[0050] (1) By constructing a causal knowledge base for the power distribution network and adopting a semi-supervised meta-learning framework, the data-driven approach is deeply integrated with physical laws. The knowledge query network can dynamically retrieve relevant causal relationships based on real-time meteorological conditions, and then correct the initial results of the main prediction network for compliance with physical laws. This ensures that the prediction results not only conform to the actual physical laws of power system operation, but also maintain high prediction accuracy, providing a reliable basis for safe power grid dispatch.

[0051] (2) By establishing a node classification mechanism based on credit scoring, monitoring nodes are innovatively divided into labeled nodes and unlabeled nodes, making full use of a small amount of high-quality labeled data and a large amount of unlabeled data for semi-supervised learning. This method significantly reduces the reliance on deploying high-precision monitoring equipment across the entire network and effectively solves the problem of insufficient monitoring data in the distribution network. At the same time, based on a multi-level security verification-based hierarchical early warning and scheduling mechanism, preventive scheduling and emergency scheduling strategies can be automatically generated according to the prediction results, realizing closed-loop intelligent decision-making from load forecasting to scheduling execution, and improving the operating efficiency and security of the distribution network. Attached Figure Description

[0052] The invention will now be further described with reference to the accompanying drawings.

[0053] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Please see Figure 1 As shown, this invention is a distribution network load forecasting and dispatching system based on multimodal data and semimodal learning, comprising:

[0056] The multimodal data acquisition and processing module is used to acquire multimodal data from multiple monitoring nodes in the distribution network in real time, including electricity load data, meteorological environment data and power grid topology data, and to classify the monitoring nodes into labeled nodes and unlabeled nodes.

[0057] The causal knowledge base construction module builds a distribution network causal knowledge base based on multimodal data with labeled nodes. The distribution network causal knowledge base includes the causal relationship between load type and meteorological conditions.

[0058] The semi-supervised meta-learning model configuration module uses the causal relationships in the causal knowledge base as prior knowledge to configure the training parameters of the semi-supervised meta-learning model, which includes a main prediction network and a knowledge query network.

[0059] The collaborative training module uses multimodal data with labeled nodes as supervision signals and multimodal data without labeled nodes as supplementary data to collaboratively train the semi-supervised meta-learning model, enabling the knowledge query network to retrieve relevant causal relationships from the causal knowledge base based on the input meteorological features.

[0060] The collaborative training module uses the trained semi-supervised meta-learning model to perform load forecasting. It inputs real-time multimodal data into the main forecasting network to obtain the initial load forecasting results, and at the same time retrieves the target causal associations that match the current meteorological conditions from the causal knowledge base through the knowledge query network.

[0061] The load forecasting and correction module, based on the retrieved target causal relationships, corrects the initial load forecast results for physical law conformity and generates the final load forecast results.

[0062] The scheduling strategy generation and execution module formulates distribution network scheduling strategies and executes corresponding load scheduling operations based on the final load forecast results.

[0063] In the multimodal data acquisition and processing module, during the power distribution network monitoring data acquisition process involved in this invention, electricity load data is collected by smart meters deployed at various monitoring points in the power distribution network, and meteorological environmental data is collected by meteorological sensors deployed in the power distribution network coverage area. These meteorological sensors include: temperature sensors, humidity sensors, precipitation sensors, wind speed sensors, and light intensity sensors. The temperature sensor is a platinum resistance temperature sensor. The humidity sensor is a capacitive humidity sensor. The precipitation sensor is a tipping bucket rain gauge. The wind speed sensor is an ultrasonic wind speed sensor. The light intensity sensor is a silicon photodiode sensor. Power grid topology data is acquired through a power grid dispatch automation system, including line connection relationships, switch status, and transformer parameters. This data is collected and stored on a minute-by-minute basis to form a complete monitoring dataset.

[0064] The credit scoring process evaluates each monitoring node based on three technical indicators. Historical data completeness rate refers to the ratio of valid data records to the required number of records within a statistical period of 30 days. Valid data records must be within a reasonable range and continuous without missing data. Abnormal data fluctuation frequency is obtained by comparing the number of times the node's load data exceeds the normal fluctuation range to the total number of data collections. The normal fluctuation range is set at ±30% of the node's historical load average. Equipment operational stability is quantitatively assessed based on the fault and maintenance records of the monitoring equipment to which the node belongs. Fault-free operation days account for 70%, and the completion rate of regular maintenance records accounts for 30%. These three indicators are normalized to a score from 0 to 100, and then weighted and summed according to a 4:3:3 weighting ratio to obtain the real-time credit score for each monitoring node.

[0065] The node tiering process sets a first threshold of 80 points. Monitoring nodes with real-time credit scores greater than or equal to 80 points are classified as labeled nodes; the data quality of these nodes meets the standard requirements and can be used in the supervised learning process. Monitoring nodes with real-time credit scores below 80 points are classified as unlabeled nodes; the data quality of these nodes does not meet the standard requirements and requires data processing and feature extraction using semi-supervised learning methods. This tiered processing method ensures the quality and reliability of the data used in subsequent learning processes.

[0066] In the causal knowledge base construction module, the candidate causal association identification process is based on historical multimodal data of labeled nodes over the past three years, employing a time-series-based causal discovery method. First, load data and meteorological parameter data are aligned to an hourly time granularity. Load types include four main categories: industrial load, commercial load, residential load, and agricultural load. Meteorological parameters include five dimensions: temperature, humidity, precipitation, wind speed, and light intensity. By calculating the lead-lag relationship between load changes and meteorological parameter changes, a time-series-based causal inference algorithm is used to analyze the mutual influence between load data and meteorological data at different time delays. For each load type and meteorological parameter combination, their mutual information entropy and conditional probability within a 24-hour time window are calculated. When the mutual information entropy exceeds a set threshold of 0.7 and the conditional probability is greater than 0.6, the combination is identified as a candidate causal association.

[0067] The multi-dimensional verification process rigorously validates the identified candidate causal relationships. Temporal consistency verification requires that load changes occur after changes in meteorological parameters, with a time delay within a reasonable range: the delay for temperature and humidity should not exceed 6 hours, the delay for precipitation should not exceed 12 hours, and the delay for wind speed and light intensity should not exceed 4 hours. The correlation strength verification calculates the Pearson correlation coefficient between load and meteorological parameters. The calculation process involves: obtaining the load data sequence and the corresponding meteorological parameter data sequence, ensuring both sequences maintain the same time scale and data length; calculating the average value of the load data and the average value of the meteorological parameter data; calculating the covariance of the load data and meteorological parameter data, i.e., the difference between the load value and its average value at each time point multiplied by the difference between the meteorological parameter value and its average value at that time point, summing all the products, and dividing by the number of data points minus 1; calculating the standard deviation of the load data sequence and the standard deviation of the meteorological parameter data sequence; and dividing the covariance by the product of the two standard deviations to obtain the Pearson correlation coefficient, which ranges from -1 to 1 and requires an absolute value greater than 0.5. Physical rationality verification is conducted by power industry experts based on their professional knowledge to ensure that the causal relationship conforms to the operating laws of the power system. Only candidate causal associations that pass all three dimensions of verification can be considered reliable causal associations.

[0068] The knowledge base construction process categorizes and stores verified reliable causal relationships according to two dimensions: load type and meteorological conditions. A graph-based storage structure is established, where nodes represent load type and meteorological conditions, and edges represent causal relationships. A confidence score is calculated for each causal relationship, weighted by a combination of time-series consistency score, correlation strength score, and physical plausibility score, with weights of 0.4, 0.4, and 0.2, respectively. Each score ranges from 0 to 1, and the final confidence score is also normalized to 0 to 1. A confidence score higher than 0.8 indicates a strong causal relationship; a score between 0.6 and 0.8 indicates a moderate causal relationship; and a score lower than 0.6 indicates a weak causal relationship. This ultimately forms a distribution network causal knowledge base containing complete causal relationships and their confidence scores.

[0069] In the semi-supervised meta-learning model configuration module, the query template library construction process is based on the causal relationships in the causal knowledge base, transforming each causal relationship into a standardized query template. Each query template contains three components: a meteorological condition input interface, a load type output interface, and a semantic description framework. The meteorological condition input interface defines the types of meteorological parameters involved in the causal relationship and their value ranges, including five dimensions: temperature, humidity, precipitation, wind speed, and light intensity. The load type output interface defines the load type classification corresponding to the causal relationship, including four types: industrial load, commercial load, residential load, and agricultural load. The semantic description framework adopts a unified natural language expression format, representing the mapping relationship from meteorological conditions to load types as a standardized description of "when [meteorological conditions], the [load type] load will undergo [change trend]". For strong causal relationships with a confidence score higher than 0.8, detailed query templates are constructed; for moderate causal relationships with a confidence score between 0.6 and 0.8, standard query templates are constructed; and for weak causal relationships with a confidence score lower than 0.6, basic query templates are constructed.

[0070] The parameter configuration table generation process determines the architecture parameters of the knowledge query network based on the structural characteristics of the query template library. The knowledge query network adopts a three-layer feedforward neural network structure. The number of input layer nodes is set to 128 based on the meteorological feature dimension, the number of hidden layer nodes is set to 64 based on the number of query templates, and the number of output layer nodes is set to 32 based on the query instruction dimension. The connection weights of each layer are initialized using the Xavier initialization method, with the initialization range calculated based on the number of input and output nodes in each layer. The feature extraction range is determined based on the meteorological parameter types defined in the query templates: temperature-related features are extracted from -40℃ to +60℃, humidity from 0% to 100%, precipitation from 0 to 500 mm, wind speed from 0 to 60 m / s, and light intensity from 0 to 2000 W / m². The parameter configuration table also defines the activation function types for each layer: the input layer uses ReLU activation, the hidden layer uses Sigmoid activation, and the output layer uses Softmax activation.

[0071] The network initialization process constructs a knowledge query network and a main prediction network based on a parameter configuration table. The initialization of the knowledge query network includes setting the number of network layers, configuring the number of nodes in each layer, initializing connection weights, and configuring activation functions. This network receives real-time meteorological feature data, processes it through three layers, and outputs the corresponding knowledge base query command. The main prediction network adopts a five-layer neural network structure. The number of nodes in the input layer is set to 256 based on the multimodal data feature dimensions. The number of nodes in the three hidden layers are set to 128, 64, and 32 respectively. The number of nodes in the output layer is set to 24 based on the prediction target, corresponding to the 24-hour load forecast value. The connection weights of each layer are initialized using the He method, the activation function is the ReLU function, and the output layer uses a linear activation function. The main prediction network is responsible for processing multimodal data composed of electricity load data, meteorological environment data, and power grid topology data, and generating initial load forecast results for the next 24 hours.

[0072] In the collaborative training module, the supervised training sample set construction process is based on historical monitoring data with labeled nodes, selecting data samples with complete records from the past three years. Each training sample contains a continuous 72-hour meteorological feature sequence as input data. The meteorological features include five dimensions: temperature, humidity, precipitation, wind speed, and light intensity. Each dimension is collected hourly to form a feature sequence at 72 time points. The corresponding load type label is used as the expected output, with load types categorized into four types: industrial load, commercial load, residential load, and agricultural load. Sample selection requires complete meteorological feature sequences and accurate load type labels, ultimately constructing a supervised training sample set containing 100,000 valid samples. The samples are divided into training, validation, and test sets in a 7:2:1 ratio to ensure the reliability and generalization ability of the training process.

[0073] The similarity matching process inputs the supervised training sample set into the knowledge query network for forward computation. After receiving the meteorological feature sequence, the knowledge query network outputs a 32-dimensional query command vector after processing through three layers. This query command vector is then matched against all causal associations stored in a causal knowledge base containing 500 verified reliable causal associations. The similarity matching employs a cosine similarity calculation method, calculating the cosine of the angle between the query command vector and each causal association representation vector to obtain a matching score within the range of 0 to 1. For each training sample, the five causal associations with the highest matching scores are selected as candidate results, and their matching scores and corresponding load type information are recorded.

[0074] The parameter optimization process constructs a multi-level similarity distribution map based on the similarity matching results. The first-level semantic feature matching degree is obtained by calculating the semantic consistency between the query command and the causal association in the load type dimension. Specifically, this is calculated by analyzing the consistency between the load type label and the load type description in the causal association, with the consistency degree determined by the hierarchical distance of the load type classification system. The second-level structural feature matching degree is obtained by comparing the structural similarity between the query command and the causal association in the meteorological condition dimension, including the matching degree of meteorological parameter types, numerical ranges, and trends. A differential loss function is calculated based on the multi-level similarity distribution map. Positive reinforcement is applied to the matching score of correct causal associations, which are those consistent with the sample's true load type label. Positive reinforcement is achieved by multiplying the matching score by a reinforcement coefficient of 1.2. Negative suppression is applied to the matching score of incorrect associations, which are causal associations inconsistent with the sample's true load type label. Negative suppression is achieved by multiplying the matching score by a suppression coefficient of 0.8. The difference amplification factor is set to 1.5 to enhance the discriminative power of the loss function for different causal associations. Gradient information is calculated based on a differentiated loss function, and the connection weights and feature extraction parameters of the knowledge query network are adjusted through backpropagation. An adaptive learning rate method is used for parameter updates, with an initial learning rate of 0.01, which decays to 0.95 times its original value after every 10,000 iterations. The update amount of connection weights is calculated based on the gradient direction and the learning rate, while the feature extraction parameters are adjusted according to the chain rule of error backpropagation. The training process continues until the decrease in the loss function value over 10 consecutive iterations is less than 0.0005, at which point the network parameters are considered to have reached a stable state. The resulting knowledge query network can accurately output query commands based on input meteorological features, effectively pointing to the correct load type associated with the input meteorological features.

[0075] In the load forecasting and correction module, the dynamic impact graph construction process is based on the target causal relationships retrieved from a knowledge query network. All causal relationships related to current meteorological conditions are obtained from the causal knowledge base, categorized into three levels: strong, moderate, and weak causal relationships. A graph structure is constructed to represent the influence relationship between load types and meteorological parameters, where nodes represent different load types and meteorological parameters, and edges represent the causal relationships between them. The weight of the edges is determined based on the confidence score of the causal relationship: edges with a confidence score between 0.8 and 1.0 have a weight of 1.0; edges with a moderate causal relationship between 0.6 and 0.8 have a weight of 0.7; and edges with a weak causal relationship between 0 and 0.6 have a weight of 0.4. Each load type node records its associated meteorological parameter type and influence direction. The influence direction is divided into positive and negative influences: a positive influence indicates that the load value increases when the meteorological parameter increases, and a negative influence indicates that the load value decreases when the meteorological parameter increases. The dynamic impact relationship diagram also records the intensity coefficient of each impact relationship. The intensity coefficient is determined based on the load change caused by changes in meteorological parameters in historical data. It is calculated by statistically analyzing the average change in load value when meteorological parameters change by one unit over the past three years.

[0076] The correction factor calculation process is based on the influence intensity of each load type in the dynamic influence relationship diagram. First, the deviation of the current meteorological parameters from historical baseline values ​​is obtained. The historical baseline values ​​are the average values ​​of meteorological parameters for the same period over the past three years. Temperature deviation is calculated by the difference between the current temperature and the historical average temperature for the same period; humidity deviation is calculated by the difference between the current humidity and the historical average humidity for the same period; precipitation deviation is calculated by the ratio of the current precipitation to the historical average precipitation for the same period; wind speed deviation is calculated by the ratio of the current wind speed to the historical average wind speed for the same period; and light intensity deviation is calculated by the ratio of the current light intensity to the historical average light intensity for the same period. Then, these deviations and the influence intensity of the corresponding load type are normalized using minimum-maximum normalization, linearly transforming the original data to the range of 0 to 1. For each load type, its sensitivity coefficient to each meteorological parameter is calculated. The sensitivity coefficient equals the influence intensity of the load type on a certain meteorological parameter multiplied by the deviation of that meteorological parameter, and then divided by the sum of the influence intensities of the load type on all meteorological parameters. Finally, the correction factor is calculated based on the sensitivity coefficient and the preset correction benchmark value. The correction benchmark value is set according to the importance of the load type: 1.2 for industrial load, 1.1 for commercial load, 1.0 for residential load, and 1.3 for agricultural load. The correction factor is calculated by multiplying the correction benchmark value by the weighted average of the sensitivity coefficient. The weights are determined based on the real-time monitoring quality of meteorological parameters, which is obtained through evaluation of the operating status of data acquisition equipment and the stability of data transmission.

[0077] The final load forecast generation process first establishes a dynamic weight allocation table based on correction factors for various load types. This table records the weight ratio of each load type under current meteorological conditions. The weight ratio is determined by the magnitude of the correction factor; a larger correction factor indicates a more significant impact of current meteorological conditions on the load type, resulting in a higher assigned weight ratio. The specific weight ratio is obtained by normalizing the correction factors for each load type using the softmax function to ensure the sum of all weight ratios is 1. Then, the initial load forecast is decomposed into forecast components for different load types. The decomposition method is based on the historical data showing the proportion of each load type in the total load, combined with real-time load characteristic analysis results. The industrial load forecast component is determined based on its typical daily load curve and current production activity index; the commercial load forecast component is determined based on its typical daily load curve and current commercial activity index; the residential load forecast component is determined based on its typical daily load curve and current residential electricity consumption characteristics; and the agricultural load forecast component is determined based on its typical daily load curve and current agricultural production seasonal characteristics. Corresponding weight adjustments are then applied to each forecast component. The weight adjustment coefficient is equal to the dynamic weight ratio of that load type divided by its original weight ratio in the initial forecast. The original weighting ratios were obtained by analyzing the average proportion of each load type under similar meteorological conditions in historical data. Finally, the weighted predicted components of each load type were recombined, taking into account the coupling relationships between load types and grid operation constraints. The coupling relationships were determined by analyzing the simultaneity rate of different load types in historical data; the simultaneity rate represents the probability that each load type will reach its peak value simultaneously. Grid operation constraints included line transmission capacity limitations, transformer load capacity limitations, and node voltage deviation limitations. The recombined prediction results also needed to undergo a rationality check. The check criteria included that the load change rate did not exceed 1.5 times the historical maximum change rate, the load curve shape conformed to the typical characteristics of that type of load, and the peak load occurrence time was within a reasonable range. The final load prediction results generated through this process maintained the accuracy of data-driven prediction while ensuring compliance with the physical laws and operational constraints of the power system.

[0078] In the load forecasting and correction module, the dynamic impact graph construction process is based on the target causal relationships retrieved from a knowledge query network. All causal relationships related to current meteorological conditions are obtained from the causal knowledge base, categorized into three levels: strong, moderate, and weak causal relationships. A graph structure is constructed to represent the influence relationship between load types and meteorological parameters, where nodes represent different load types and meteorological parameters, and edges represent the causal relationships between them. The weight of the edges is determined based on the confidence score of the causal relationship: edges with a confidence score between 0.8 and 1.0 have a weight of 1.0; edges with a moderate causal relationship between 0.6 and 0.8 have a weight of 0.7; and edges with a weak causal relationship between 0 and 0.6 have a weight of 0.4. Each load type node records its associated meteorological parameter type and influence direction. The influence direction is divided into positive and negative influences: a positive influence indicates that the load value increases when the meteorological parameter increases, and a negative influence indicates that the load value decreases when the meteorological parameter increases. The dynamic impact relationship diagram also records the intensity coefficient of each impact relationship. The intensity coefficient is determined based on the load change caused by changes in meteorological parameters in historical data. It is calculated by statistically analyzing the average change in load value when meteorological parameters change by one unit over the past three years.

[0079] The correction factor calculation process is based on the influence intensity of each load type in the dynamic influence relationship diagram. First, the deviation of the current meteorological parameters from historical baseline values ​​is obtained. The historical baseline values ​​are the average values ​​of meteorological parameters for the same period over the past three years. Temperature deviation is calculated by the difference between the current temperature and the historical average temperature for the same period; humidity deviation is calculated by the difference between the current humidity and the historical average humidity for the same period; precipitation deviation is calculated by the ratio of the current precipitation to the historical average precipitation for the same period; wind speed deviation is calculated by the ratio of the current wind speed to the historical average wind speed for the same period; and light intensity deviation is calculated by the ratio of the current light intensity to the historical average light intensity for the same period. Then, these deviations and the influence intensity of the corresponding load type are normalized using minimum-maximum normalization, linearly transforming the original data to the range of 0 to 1. For each load type, its sensitivity coefficient to each meteorological parameter is calculated. The sensitivity coefficient equals the influence intensity of the load type on a certain meteorological parameter multiplied by the deviation of that meteorological parameter, and then divided by the sum of the influence intensities of the load type on all meteorological parameters. Finally, the correction factor is calculated based on the sensitivity coefficient and the preset correction benchmark value. The correction benchmark value is set according to the importance of the load type: 1.2 for industrial load, 1.1 for commercial load, 1.0 for residential load, and 1.3 for agricultural load. The correction factor is calculated by multiplying the correction benchmark value by the weighted average of the sensitivity coefficient. The weights are determined based on the real-time monitoring quality of meteorological parameters, which is obtained through evaluation of the operating status of data acquisition equipment and the stability of data transmission.

[0080] The final load forecast generation process first establishes a dynamic weight allocation table based on correction factors for various load types. This table records the weight ratio of each load type under current meteorological conditions. The weight ratio is determined by the magnitude of the correction factor; a larger correction factor indicates a more significant impact of current meteorological conditions on the load type, resulting in a higher assigned weight ratio. The specific weight ratio is obtained by normalizing the correction factors for each load type using the softmax function to ensure the sum of all weight ratios is 1. Then, the initial load forecast is decomposed into forecast components for different load types. The decomposition method is based on the historical data showing the proportion of each load type in the total load, combined with real-time load characteristic analysis results. The industrial load forecast component is determined based on its typical daily load curve and current production activity index; the commercial load forecast component is determined based on its typical daily load curve and current commercial activity index; the residential load forecast component is determined based on its typical daily load curve and current residential electricity consumption characteristics; and the agricultural load forecast component is determined based on its typical daily load curve and current agricultural production seasonal characteristics. Corresponding weight adjustments are then applied to each forecast component. The weight adjustment coefficient is equal to the dynamic weight ratio of that load type divided by its original weight ratio in the initial forecast. The original weighting ratios were obtained by analyzing the average proportion of each load type under similar meteorological conditions in historical data. Finally, the weighted predicted components of each load type were recombined, taking into account the coupling relationships between load types and grid operation constraints. The coupling relationships were determined by analyzing the simultaneity rate of different load types in historical data; the simultaneity rate represents the probability that each load type will reach its peak value simultaneously. Grid operation constraints included line transmission capacity limitations, transformer load capacity limitations, and node voltage deviation limitations. The recombined prediction results also needed to undergo a rationality check. The check criteria included that the load change rate did not exceed 1.5 times the historical maximum change rate, the load curve shape conformed to the typical characteristics of that type of load, and the peak load occurrence time was within a reasonable range. The final load prediction results generated through this process maintained the accuracy of data-driven prediction while ensuring compliance with the physical laws and operational constraints of the power system.

[0081] In the scheduling strategy generation and execution module, a multi-level safety verification process is conducted based on the final load forecast results. Line load rate assessment is obtained by calculating the ratio of the predicted load to the rated current carrying capacity of each line. When the ratio exceeds 80%, it is classified as a heavily loaded line; when it exceeds 100%, it is classified as an overloaded line. Transformer capacity margin analysis is obtained by calculating the difference between the predicted load and the rated capacity of each transformer. When the difference is less than 20% of the transformer's rated capacity, it is classified as a capacity shortage state; when the difference is less than 0, it is classified as an overload state. Node voltage stability calculation analyzes the voltage deviation of each node through power flow calculations. When the voltage deviation exceeds ±5% of the rated voltage, it is classified as a voltage unstable node; when it exceeds ±7%, it is classified as a node with severely exceeded voltage limits. The above verifications are performed every 15 minutes, updating the verification conclusions with the latest load forecast results.

[0082] The generation process of tiered early warning signals is automatically triggered based on safety verification results. The first threshold is set at 85% of the equipment's rated capacity. When the predicted load exceeds the first threshold, a yellow warning signal is generated, and a preventative dispatching procedure is initiated. The second threshold is set at 95% of the equipment's rated capacity. When the predicted load exceeds the second threshold, a red warning signal is generated, and an emergency dispatching procedure is initiated. The warning signal includes three elements: the affected equipment identifier, the degree of limit violation, and the expected duration. The warning level is dynamically adjusted based on the number of simultaneously alarming devices. When more than three devices in the same power supply area alarm simultaneously, the warning level is automatically upgraded by one level. Warning information is sent to relevant operating personnel through the dispatch automation system and recorded in the operation log.

[0083] The dispatching operation execution process implements corresponding measures according to the warning level. Preventative dispatching operations include adjusting transformer tap positions, with each adjustment increment at 1.25% of the rated voltage, and the cumulative adjustment within 24 hours not exceeding 5% of the rated voltage; switching reactive power compensation devices, determining the switching capacity based on the voltage deviation, with each operation not exceeding 25% of the total device capacity. Emergency dispatching operations include activating the load shedding plan, prioritizing the shedding of loads at importance level three according to pre-determined load importance classifications, with a single shedding amount not exceeding 10% of the total load; adjusting distributed power generation output, prioritizing photovoltaic power generation output based on power type and operating status, followed by wind power generation output, with the adjustment increment determined according to actual needs, and a maximum adjustment of 50% of the current output. After all dispatching operations are executed, the effect must be confirmed within 5 minutes. If the problem is not resolved, the backup dispatching plan is activated. The entire dispatching operation process is recorded and an execution report is generated, including information such as operation time, operation content, and operation effect.

[0084] The working principle of this invention is as follows: Electricity load data and meteorological environment data are collected by deploying smart meters and meteorological sensors. Combined with power grid topology data obtained from the power grid dispatch automation system, a multimodal dataset containing labeled and unlabeled nodes is constructed. A distribution network causal knowledge base is established based on the labeled node data. Reliable causal relationships between load types and meteorological conditions are extracted through time-series relationship analysis and multi-dimensional verification. Causal relationships are used as prior knowledge to configure a semi-supervised meta-learning model. Through collaborative training, the knowledge query network can retrieve relevant causal relationships based on meteorological characteristics. In the prediction stage, the initial load prediction from the main prediction network and the target causal relationships retrieved by the knowledge query network are combined. Dynamic influence relationship diagrams and correction factors are used to calculate physical law compliance corrections. Finally, multi-level safety checks are performed based on the corrected load prediction results. Preventive or emergency dispatch operations are executed according to the warning level, forming a complete technical closed loop from data acquisition to dispatch execution.

[0085] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A power distribution network load forecasting and dispatching system based on multi-modal data and semi-modal learning, characterized in that, include: The multimodal data acquisition and processing module is used to acquire multimodal data from multiple monitoring nodes in the power distribution network in real time and divide the monitoring nodes into labeled nodes and unlabeled nodes. The causal knowledge base construction module establishes a causal knowledge base for the distribution network based on multimodal data with labeled nodes, which includes the causal relationship between load type and meteorological conditions. The semi-supervised meta-learning model configuration module uses causal relationships as prior knowledge to configure a semi-supervised meta-learning model that includes a main prediction network and a knowledge query network. The collaborative training module uses multimodal data with labeled nodes as supervision signals and multimodal data without labeled nodes as supplementary data to train the knowledge query network to retrieve relevant causal relationships from the causal knowledge base. The load forecasting and correction module generates an initial load forecast through the main forecasting network, while using a knowledge query network to retrieve the target causal relationships under the current meteorological conditions, and performs physical law conformity correction on the initial load forecast. The scheduling strategy generation and execution module formulates and executes the distribution network scheduling strategy based on the final load forecast results.

2. The distribution network load forecasting and dispatching system based on multimodal data and semimodal learning according to claim 1, characterized in that, The process of dividing monitoring nodes into tagged nodes and untagged nodes specifically includes: Based on the historical data integrity rate, data anomaly fluctuation frequency, and equipment operation stability indicators of each monitoring node, a real-time credit score is calculated for each monitoring node. Based on the real-time credit score, the monitoring nodes are classified and processed. Monitoring nodes with a real-time credit score greater than or equal to the first threshold are automatically classified as labeled nodes, and monitoring nodes with a real-time credit score less than the first threshold are classified as unlabeled nodes.

3. The distribution network load forecasting and dispatching system based on multimodal data and semimodal learning according to claim 1, characterized in that, The establishment of a causal knowledge base for the distribution network that includes the causal relationship between load type and meteorological conditions specifically includes: Based on historical multimodal data with labeled nodes, candidate causal relationships between load types and meteorological conditions are identified, with potential causal relationships discovered by analyzing the temporal relationship and statistical dependence between load changes and meteorological parameter changes. The candidate causal associations were verified in multiple dimensions, including verifying the temporal consistency, association strength and physical rationality between load changes and meteorological parameter changes, and screening out the reliable causal associations that passed the verification. The verified reliable causal relationships are classified and stored according to load type and meteorological conditions to establish a structured causal knowledge base. The confidence score of each causal relationship is calculated to form a complete causal knowledge base for the power distribution network.

4. The distribution network load forecasting and dispatching system based on multimodal data and semimodal learning according to claim 1, characterized in that, The method of using causal relationships as prior knowledge and configuring a semi-supervised meta-learning model comprising a main prediction network and a knowledge query network specifically includes: Based on the causal relationships in the causal knowledge base, a query template library is constructed, where each query template corresponds to a mapping relationship from a type of meteorological conditions to a type of load and contains a complete semantic description framework; Based on the structural characteristics of the query template library, a parameter configuration table for the knowledge query network is generated. The parameter configuration table defines the connection weight initialization scheme and feature extraction range for each layer of the knowledge query network. The component structures of the knowledge query network and the main forecast network are initialized based on the parameter configuration table. The knowledge query network is responsible for receiving meteorological features and outputting knowledge base query instructions, while the main forecast network is responsible for processing multimodal data and generating load forecasts.

5. A distribution network load forecasting and dispatching system based on multimodal data and semimodal learning according to claim 1, characterized in that, The training knowledge query network retrieves relevant causal relationships from the causal knowledge base, specifically including: A supervised training sample set is constructed based on the meteorological feature sequences of labeled nodes and their corresponding load type labels; The supervised training sample set is input into the knowledge query network for forward computation to obtain the query instructions output by the knowledge query network, and the query instructions are matched with the actual causal relationships in the causal knowledge base for similarity. The loss function is calculated based on the similarity matching results, and the parameters of the knowledge query network are adjusted through backpropagation so that the query instructions output by the knowledge query network can accurately point to the load type associated with the input meteorological characteristics.

6. A distribution network load forecasting and dispatching system based on multimodal data and semimodal learning according to claim 5, characterized in that, The step of calculating the loss function based on the similarity matching results and adjusting the parameters of the knowledge query network through backpropagation specifically includes: Based on the matching scores between query commands and various causal relationships in the causal knowledge base, a multi-level similarity distribution map is constructed, where the first level measures the degree of matching of semantic features and the second level evaluates the degree of matching of structural features. The differential loss function is calculated based on the multi-level similarity distribution map. Positive reinforcement is applied to the matching score of correct causal associations, negative suppression is applied to the matching score of incorrect associations, and the discriminativeness of the loss function is enhanced by the differential amplification factor. Gradient information is calculated based on the differential loss function, and the connection weights and feature extraction parameters of the knowledge query network are adjusted through the backpropagation process.

7. A distribution network load forecasting and dispatching system based on multimodal data and semimodal learning according to claim 1, characterized in that, The physical conformity correction of the initial load forecast specifically includes: Based on the target causal relationships obtained from knowledge query network retrieval, a dynamic influence relationship diagram between load type and meteorological parameters is constructed. Based on the influence intensity of each load type in the dynamic influence relationship diagram, calculate the correction factors for each type of load under the current meteorological conditions; The correction factor is weighted and fused with the initial load forecast generated by the main forecast network to generate a final load forecast that conforms to physical laws.

8. A distribution network load forecasting and dispatching system based on multimodal data and semimodal learning according to claim 7, characterized in that, The process of obtaining the correction factor is as follows: Based on the influence intensity of each load type in the dynamic influence relationship diagram, the degree of deviation of the current meteorological parameters from the historical baseline value is obtained; The degree of deviation and the influence intensity of the corresponding load type are normalized to obtain the sensitivity coefficient of each load type under the current meteorological conditions; Based on the sensitivity coefficient and the preset correction benchmark value, the correction factor for various types of loads is calculated through weighted product.

9. A distribution network load forecasting and dispatching system based on multimodal data and semimodal learning according to claim 7, characterized in that, The generation of the final load forecast that conforms to physical laws specifically includes: A dynamic weight allocation table is established based on the correction factors for various types of loads, and the weight ratio of various types of loads in the final forecast is determined according to the magnitude of the correction factors. Based on the weight ratios in the dynamic weight allocation table, the initial load forecast is decomposed into forecast components of different load types, and corresponding weight adjustments are made to each forecast component. The weighted load type forecast components are recombined to generate a final load forecast that conforms to physical laws.

10. A distribution network load forecasting and dispatching system based on multimodal data and semimodal learning according to claim 1, characterized in that, The process of formulating and executing distribution network dispatch strategies based on final load forecast results specifically includes: Multi-level safety verification of power grid operation status is carried out based on the final load forecast results, including line load rate assessment, transformer capacity margin analysis and node voltage stability calculation. Based on the safety verification results, a graded early warning signal is generated. When the predicted load exceeds the first threshold of the equipment's rated capacity, preventive scheduling is initiated. When it exceeds the second threshold, emergency scheduling is initiated. Perform dispatching operations corresponding to the warning level. Preventive dispatching includes adjusting transformer taps and switching reactive power compensation devices, while emergency dispatching includes activating load shedding plans and adjusting the output of distributed power sources.

Images