Urban power distribution method and system based on load prediction

Abstract

The invention belongs to the technical field of urban power distribution, and provides an urban power distribution method and system based on load prediction, and the method comprises the steps: collecting power load data, calculating sequence gradient dynamic adjustment noise injection to generate desensitization data, carrying out closed-loop control to accumulate privacy budget, and aggregating the desensitization data through edge nodes. Training a local load prediction model by utilizing federated learning and knowledge distillation, summarizing and updating a global model at a cloud end, fusing multi-source data, predicting and generating a power distribution strategy by utilizing an interpretable model, comparing actual and predicted data, dynamically adjusting model parameters and privacy budget, and forming closed-loop optimization; the system correspondingly comprises a privacy perception data preprocessing module, an edge federal knowledge distillation training module, a cloud interpretability prediction decision module and a closed loop feedback optimization module, data privacy is protected, load prediction precision and power distribution accuracy are improved, system robustness is enhanced, and the method is suitable for urban power distribution.

Description

Technical Field

[0001] This invention belongs to the field of urban power distribution technology, specifically to an urban power distribution method and system based on load forecasting. Background Technology

[0002] In urban power distribution, the application of high-frequency data from smart meters can provide rich information for load forecasting, but the issue of protecting the privacy of power load data also becomes prominent.

[0003] Existing technologies such as generalization and suppression often sacrifice the availability of power load data while protecting its privacy, leading to a decrease in the accuracy of power load data predicted by the model.

[0004] In addition, traditional power distribution methods have shortcomings in data processing, model optimization, and decision transparency, making it difficult to meet the requirements of urban power systems for real-time performance, security, and reliability.

[0005] To this end, the present invention provides a method and system for urban power distribution based on load forecasting. Summary of the Invention

[0006] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0007] The technical solution adopted by this invention to solve its technical problem is: In a first aspect, the present invention provides a method for urban power distribution based on load forecasting, comprising: S1: Collect power load data from various areas of the city through smart meters, perform sequence gradient calculation on the collected power load data, and dynamically adjust the noise injection amplitude according to the sequence gradient to inject Laplace noise into the power load data value to generate desensitized power load data. S2: Edge nodes collect and aggregate desensitized power load data from each region, and use a federated learning and knowledge distillation fusion model to train the local load prediction model. First, a global model and an edge model are defined. Then, the knowledge of the global model is transferred to the local prediction model through federated knowledge distillation, thereby improving the performance of the local prediction model without sharing the original power load data. S3: The cloud receives the local prediction model parameters trained by each edge node, summarizes and updates the global model, combines multi-source data, uses the interpretability-enhanced load prediction model to output power load prediction data, and generates power allocation strategies based on the power load prediction data. S4: Compare the actual power load data with the power load forecast data output by the load forecast model, and dynamically adjust the parameters of the local forecast model based on the comparison results to optimize the accuracy of the power load forecast data and power allocation strategy.

[0008] As a further improvement of the present invention, the process of collecting power load data of various urban areas through smart meters is as follows: Electricity load data from various areas of the city is collected using smart meters to form a raw electricity load data sequence. ,in, This represents the load value at the nth time point.

[0009] As a further improvement of the present invention, the specific process of performing sequence gradient calculation is as follows: Calculate the time series gradient of the power load data to capture the rate of load change. For each time point t of the power load data, calculate the sequence gradient. Defined as: ,in, The gradient represents the magnitude of load change at adjacent time points. The power load data value at time point t. For the time point is The larger the absolute value of the sequence gradient of the power load data, the more severe the load fluctuation.

[0010] As a further improvement of the present invention, the specific process of dynamically adjusting the noise injection amplitude according to the sequence gradient is as follows: Based on sequence gradient The noise injection amplitude is dynamically adjusted, specifically during periods of high change rate. Injecting less noise preserves more data details, especially during periods of low change rate. Injecting significant noise enhances privacy protection; noise amplitude The adjustment formula is: ,in, For the initial privacy budget, The adjustable parameters are used to control the sensitivity of noise to changes with the gradient; noise amplitude. The adjustment formula uses the sigmoid function to achieve a nonlinear mapping between the sequence gradient and noise, ensuring that high sequence gradients are avoided. near When the sequence gradient is low Approaching 0.

[0011] As a further improvement of the present invention, the specific process of injecting Laplace noise into the power load data value to generate desensitized power load data is as follows: For each power load data value Injecting Laplace noise Generate desensitized data : Among them, Laplace noise Follows a Laplace distribution: ,in, The sensitivity parameter is usually set to 1, which represents the maximum magnitude of data variation. This is a dynamically adjusted privacy budget.

[0012] As a further improvement of the present invention, the specific process by which the edge node collects and aggregates the desensitized power load data in each region is as follows: Edge nodes collect anonymized power load data from each region. ,in, This represents the desensitized load value at the nth time point; the desensitized power load data collected from each region by the edge nodes are summed to form a regional load dataset.

[0013] As a further improvement of the present invention, the specific process of training the local load prediction model using the federated learning and knowledge distillation fusion model, first defining the global model and the edge model, is as follows: Define the global model as: denoted as It is a high-performance, complex model maintained in the cloud, possessing knowledge of global data training; Define the edge model as: denoted as It is a lightweight model deployed locally on edge nodes.

[0014] As a further improvement of the present invention, the specific process by which the cloud receives the local prediction model parameters trained by each edge node and summarizes and updates the global model is as follows: The cloud receives the local prediction model parameters trained by each edge node. Where m is the number of edge nodes, the global model parameters are updated through a federated learning aggregation strategy. The specific process is as follows: Weighted aggregation of each edge model based on regional load scale: ,in, Let i be the amount of load data covered by the i-th edge node. This represents the total global data volume. These are the parameters for the i-th local prediction model.

[0015] As a further improvement of the present invention, the multi-source data specifically includes: Meteorological data, economic indicators, and historical power load data; among which, meteorological data includes: temperature, humidity, and sunshine; economic indicators include: GDP growth rate and industrial output value.

[0016] In a second aspect, the present invention provides an urban power distribution system based on load forecasting, comprising: Privacy-aware data preprocessing module: Collects power load data from various areas of the city through smart meters, performs sequence gradient calculation on the collected power load data, dynamically adjusts the noise injection amplitude according to the sequence gradient, and injects Laplace noise into the power load data values ​​to generate desensitized power load data; Edge Federated Knowledge Distillation Training Module: Edge nodes collect and aggregate desensitized power load data from various regions, and use a federated learning and knowledge distillation fusion model to train the local load prediction model. First, a global model and an edge model are defined. Then, the knowledge of the global model is transferred to the local prediction model through federated knowledge distillation, which improves the performance of the local prediction model without sharing the original power load data. The cloud-based interpretable prediction decision module receives the local prediction model parameters trained by each edge node, summarizes and updates the global model, combines multi-source data, uses the interpretable-enhanced load prediction model to output power load prediction data, and generates power allocation strategies based on the power load prediction data. Closed-loop feedback optimization module: Compares the actual power load data with the power load forecast data output by the load forecast model, and dynamically adjusts the parameters of the local forecast model based on the comparison results to optimize the accuracy of the power load forecast data and power allocation strategy.

[0017] The beneficial effects of this invention are as follows: 1. By dynamically adjusting the noise injection amplitude through sequence gradients, features are retained with less noise during periods of high volatility and enhanced privacy during periods of low volatility. Combined with cumulative privacy budget closed-loop control, global privacy leakage is avoided, while key features such as load mutations are preserved to maintain temporal continuity and reduce prediction errors during periods of high volatility. A federated learning and knowledge distillation fusion model is used, with local training at edge nodes combined with global knowledge transfer in the cloud, which can improve the performance of local models without sharing original data. Multi-source data fusion and interpretable attention mechanisms quantify feature importance and enhance prediction robustness.

[0018] 2. Based on the prediction data generation and allocation strategy, taking into account grid transmission capacity constraints, power generation costs, and low-carbon orientation, multi-objective optimization is achieved through algorithms such as linear programming to ensure power supply reliability and economy. It is adaptable to different scenarios such as residential and industrial applications. The closed-loop feedback mechanism compares actual load with prediction data to dynamically adjust model parameters and privacy budget, forming a closed loop from prediction to decision to optimization. It adapts to changes in grid topology and fluctuations in new energy sources, improving system stability. The Sigmoid function achieves nonlinear noise control and implements differential privacy theory. The edge-cloud collaborative architecture combined with an interpretable model combines data security, computational efficiency, and decision transparency, meeting the requirements of power industry standards and regulations. Attached Figure Description

[0019] The invention will now be further described with reference to the accompanying drawings.

[0020] Figure 1 This is a flowchart of the steps of the urban power distribution method based on load prediction of the present invention; Figure 2 This is a system module diagram of the urban power distribution system based on load prediction of the present invention. Detailed Implementation

[0021] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0022] Example 1 like Figure 1 As shown in the embodiment of the present invention, the urban power distribution method based on load forecasting includes: S1: Collect power load data from various areas of the city through smart meters, perform sequence gradient calculation on the collected power load data, and dynamically adjust the noise injection amplitude according to the sequence gradient to inject Laplace noise into the power load data value to generate desensitized power load data. The process of collecting power load data from various areas of the city through smart meters is as follows: Electricity load data from various areas of the city is collected using smart meters to form a raw electricity load data sequence. ,in, This represents the load value at the nth time point; Sequence gradient calculation is performed on the collected power load data; The specific process for calculating the sequence gradient is as follows: Calculate the time series gradient of the power load data to capture the rate of load change. For each time point t of the power load data, calculate the sequence gradient. Defined as: ,in, The gradient represents the magnitude of load change at adjacent time points. The power load data value at time point t. For the time point is The larger the absolute value of the sequence gradient of the power load data, the more severe the load fluctuation. Based on the calculated sequence gradient, the noise injection amplitude is dynamically adjusted. The specific process of dynamically adjusting the noise injection amplitude based on the sequence gradient is as follows: Based on sequence gradient The noise injection amplitude is dynamically adjusted, specifically during periods of high change rate. Injecting less noise preserves more data details, especially during periods of low change rate. Injecting significant noise enhances privacy protection; noise amplitude The adjustment formula is: ,in, For the initial privacy budget, The adjustable parameters are used to control the sensitivity of noise to changes with the gradient; noise amplitude. The adjustment formula uses the sigmoid function to achieve a nonlinear mapping between the sequence gradient and noise, ensuring that high sequence gradients are avoided. near When the sequence gradient is low Approaching 0; The specific process of injecting Laplace noise into the power load data value to generate desensitized power load data is as follows: For each power load data value Injecting Laplace noise Generate desensitized data : Among them, Laplace noise Follows a Laplace distribution: ,in, The sensitivity parameter is usually set to 1, which represents the maximum magnitude of data variation. For dynamically adjusted privacy budgets; If only a single power load data value is considered Dynamic adjustments could lead to an uncontrollable risk of global privacy breaches, such as the cumulative effect of small privacy budgets at a large number of time points exceeding the security threshold. Therefore, closed-loop control of the cumulative privacy budget throughout the preprocessing process is necessary to ensure that the cumulative privacy budget does not exceed a preset threshold. The specific formula for calculating the cumulative privacy budget is as follows: ,in, For dynamic privacy budgets at a single point in time, Accumulate a global privacy budget; The calculated global cumulative privacy budget is compared with the preset threshold. Compare; like Then the noise injection continues; like If noise injection is stopped, the balance between privacy protection and data availability will be maintained. It should be noted that the preset threshold is a reference value set by technical personnel in this industry based on production experience, and is used as a reference value to measure noise injection and terminate noise injection. By distinguishing high and low fluctuation periods through sequence gradients, dynamically adjusting noise, and combining cumulative privacy budget closed loop, the risk of leakage due to multiple overlapping periods is prevented, key features of power load data mutations are preserved, time series continuity is maintained, and prediction accuracy is improved. The nonlinear control of noise is achieved by using the Sigmoid function, implementing differential privacy theory, and improving the robustness of the system. S2: Edge nodes collect and aggregate desensitized power load data from each region, and use a federated learning and knowledge distillation fusion model to train the local load prediction model. First, a global model and an edge model are defined. Then, the knowledge of the global model is transferred to the local prediction model through federated knowledge distillation, thereby improving the performance of the local prediction model without sharing the original power load data. The specific process by which the edge nodes collect and aggregate the de-identified power load data from each region is as follows: Edge nodes collect anonymized power load data from each region. ,in, This represents the anonymized load value at time point n; the anonymized power load data collected from each region by the edge nodes are summed to form a regional load dataset. ; The specific process of training a local load prediction model using a fusion model of federated learning and knowledge distillation, first defining the global model and the edge model, is as follows: Define the global model as: denoted as It is a high-performance complex model maintained in the cloud, which has the knowledge of global data training. The high-performance complex model includes, but is not limited to: LSTM and Transformer models. Define the edge model as: denoted as It is a lightweight model deployed locally on edge nodes, with a simpler structure than the global model, including but not limited to: shallow neural networks; The specific process of transferring knowledge from the global model to the local model through federated knowledge distillation is as follows: Global model knowledge extraction cloud will extract the global model's knowledge. The intermediate layer output is transmitted to the edge nodes. The intermediate layer output includes, but is not limited to: hidden layer features, soft tags, and power load data input to the edge nodes. The output of the global model is a soft label. ,in, It is a vector containing the probabilities of each predicted class. This is the set of parameters for the global model. The power load data input to the edge nodes. For global models; Using forward propagation of the edge model, edge nodes use regional load datasets. Input to edge model , get output ,in, For edge models, For data in the regional load dataset, This is the parameter set for the edge model; The difference between the outputs of the marginal model and the global model is calculated as the distillation loss. The formula for calculating distillation loss is: ,in, and , respectively, represent the predicted probabilities of the global model and the marginal model for the i-th class; The calculation of distillation loss in local supervised learning combines the real label y of local data to calculate the traditional supervised learning loss. : Where m is the number of local training samples, and Let be the input and the actual load value of the j-th sample, respectively. This is the parameter set for the edge model; By combining distillation loss and supervised learning loss, a joint loss function is formed. Update the edge model parameters using gradient descent. : ,in, Weighting coefficients This is used to balance the importance of knowledge distillation and supervised learning loss. For the joint loss function, To monitor learning loss, This is due to distillation losses; For regional load datasets Perform multiple rounds of iterative training until the local prediction model converges or reaches the preset number of iterations, ultimately obtaining the optimized local prediction model. ; By aggregating de-identified data at edge nodes and training local models using federated learning and knowledge distillation, the benefits include: regional data aggregation reduces data transmission volume and protects user privacy; federated learning avoids sharing of raw data, meeting privacy protection requirements; knowledge distillation transfers global model knowledge from the cloud to the edge, improving the generalization ability of local models; lightweight edge models adapt to the computing power limitations of edge nodes, reducing computational overhead; and the joint optimization of distillation loss and supervised learning loss balances global knowledge transfer and local data feature learning, improving prediction accuracy.

[0023] S3: The cloud receives the local prediction model parameters trained by each edge node, summarizes and updates the global model, combines multi-source data, uses the interpretability-enhanced load prediction model to output power load prediction data, and generates power allocation strategies based on the power load prediction data. The specific process by which the cloud receives the local prediction model parameters trained by each edge node and summarizes and updates the global model is as follows: The cloud receives the local prediction model parameters trained by each edge node. Where m is the number of edge nodes, the global model parameters are updated through a federated learning aggregation strategy. The specific process is as follows: Weighted aggregation of each edge model based on regional load scale: ,in, Let i be the amount of load data covered by the i-th edge node. This represents the total global data volume. These are the parameters of the i-th local prediction model; The multi-source data specifically includes: meteorological data, economic indicators, and historical power load data; among which, meteorological data includes: temperature, humidity, and light intensity; economic indicators include: GDP growth rate and industrial output value. The specific process of outputting power load forecast data using the load forecasting model with enhanced interpretability is as follows: Integrated meteorological data: temperature ,humidity ,illumination Economic indicators: GDP growth rate Industrial output value and historical power load data Construct the input feature vector: Where k is the length of the historical power load data window; The model structure employing an attention mechanism combined with feature importance analysis involves the following steps: For the input feature vector Attention weights are generated along the time dimension to quantify the contribution of each time step to the prediction; attention scores are also used. The formula for calculation is: Where W, b, and w are the local prediction model parameters, and T is the time series length. The input feature vector at time step t, The hyperbolic tangent activation function is used. It is an exponential function; Gradient-weighted class activation mapping is used to calculate the importance scores of features in the feature vector, such as meteorological data and economic indicators, to the prediction results. The importance score is calculated using the following formula: ,in Let M be the j-th dimension feature at time point t, and M be the number of samples. For power load forecasting data; The power load forecast data is obtained through the above calculations. ; Based on power load forecast data, the specific process for generating the power allocation strategy is as follows: The fusion model outputs power load forecast data for various urban areas for future periods. ,in, Using the predicted power load data for the i-th region in the future time period, and combining grid transmission capacity constraints and generation cost factors, an optimization algorithm is used to generate a power allocation strategy. ,in, For the power supply of region i, the power distribution strategy satisfies: ,in, Total power supply The allowable load fluctuation safety factor; By aggregating edge model parameters in the cloud and combining them with multi-source data, high-precision load forecasting is achieved using interpretable models. At the same time, optimized power allocation strategies are generated. Its beneficial effects include: federated aggregation improves the generalization ability of the global model, multi-source data fusion enhances the robustness of prediction, attention mechanism and feature importance analysis improve the interpretability of prediction, and the generation of economical and reliable allocation strategies by combining grid constraints and cost factors, adapting to the power dispatching needs of multiple scenarios.

[0024] S4: Compare the actual power load data with the power load forecast data output by the load forecast model, and dynamically adjust the parameters of the local forecast model based on the comparison results to optimize the accuracy of the power load forecast data and power allocation strategy. The specific process of comparing the actual power load data with the power load forecast data output by the load forecasting model is as follows: Actual power load data With power load forecast data Compare and calculate the prediction error at each time point. Prediction error The calculation formula is: ,in, For power load forecasting data, This is actual power load data; The specific process of dynamically adjusting the parameters of the local prediction model based on the comparison results is as follows: Based on prediction error Dynamically adjust the privacy budget of the differential privacy algorithm ; like This indicates insufficient data availability; therefore, reducing noise injection is necessary, i.e., increasing... ; like This indicates that privacy protection is sufficient, and increasing noise injection reduces... ; It should be noted that the error threshold is a reference value set by technical personnel in this industry, used to dynamically balance data availability and the strength of privacy protection; The prediction error is fed back to the edge nodes to adjust the local prediction model parameters. The gradient descent method is used to adjust the global model aggregation weights. The adjusted privacy budget and model parameters are applied to the next round of data preprocessing and prediction, forming a closed loop from prediction to error analysis to parameter adjustment to re-prediction, until the error converges to the preset range.

[0025] By comparing actual and predicted load data, the privacy budget and model parameters are dynamically adjusted to form a closed-loop optimization. Its beneficial effects include: improving prediction accuracy by dynamically balancing privacy protection and data availability based on errors; enhancing adaptability to dynamic changes in the power grid by optimizing model parameters through gradient descent; and ensuring system robustness by continuously iterating prediction and allocation strategies through the closed-loop mechanism, ensuring that errors converge to a preset range.

[0026] The technical solution of this invention is as follows: Electricity load data from various urban areas is collected using smart meters. Sequence gradient calculation is performed on the collected electricity load data, and the noise injection amplitude is dynamically adjusted based on the sequence gradient. Laplace noise is injected into the electricity load data values ​​to generate desensitized electricity load data. Edge nodes collect and aggregate the desensitized electricity load data from each area. A local load prediction model is trained using a federated learning and knowledge distillation fusion model. First, a global model and an edge model are defined. Knowledge from the global model is transferred to the local prediction model through federated knowledge distillation, improving the performance of the local prediction model without sharing the original electricity load data. The cloud receives the parameters of the local prediction model trained by each edge node, summarizes and updates the global model, combines multi-source data, and outputs electricity load prediction data using an interpretability-enhanced load prediction model. An electricity allocation strategy is generated based on the electricity load prediction data. The actual electricity load data is compared with the electricity load prediction data output by the load prediction model. Based on the comparison results, the parameters of the local prediction model are dynamically adjusted to optimize the accuracy of the electricity load prediction data and the electricity allocation strategy.

[0027] Example 2 like Figure 2 As shown in Embodiment 1, the present invention provides an urban power distribution system based on load forecasting, comprising: Privacy-aware data preprocessing module: Collects power load data from various areas of the city through smart meters, performs sequence gradient calculation on the collected power load data, dynamically adjusts the noise injection amplitude according to the sequence gradient, and injects Laplace noise into the power load data values ​​to generate desensitized power load data; Edge Federated Knowledge Distillation Training Module: Edge nodes collect and aggregate desensitized power load data from various regions, and use a federated learning and knowledge distillation fusion model to train the local load prediction model. First, a global model and an edge model are defined. Then, the knowledge of the global model is transferred to the local prediction model through federated knowledge distillation, which improves the performance of the local prediction model without sharing the original power load data. The cloud-based interpretable prediction decision module receives the local prediction model parameters trained by each edge node, summarizes and updates the global model, combines multi-source data, uses the interpretable-enhanced load prediction model to output power load prediction data, and generates power allocation strategies based on the power load prediction data. Closed-loop feedback optimization module: Compares the actual power load data with the power load forecast data output by the load forecast model, and dynamically adjusts the parameters of the local forecast model based on the comparison results to optimize the accuracy of the power load forecast data and power allocation strategy.

[0028] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for urban power distribution based on load forecasting, characterized in that: include: S1: Collect power load data from various areas of the city through smart meters, perform sequence gradient calculation on the collected power load data, and dynamically adjust the noise injection amplitude according to the sequence gradient to inject Laplace noise into the power load data value to generate desensitized power load data. S2: Edge nodes collect and aggregate desensitized power load data from each region, and use a federated learning and knowledge distillation fusion model to train the local load prediction model. First, a global model and an edge model are defined. Then, the knowledge of the global model is transferred to the local prediction model through federated knowledge distillation, thereby improving the performance of the local prediction model without sharing the original power load data. S3: The cloud receives the local prediction model parameters trained by each edge node, summarizes and updates the global model, combines multi-source data, uses the interpretability-enhanced load prediction model to output power load prediction data, and generates power allocation strategies based on the power load prediction data. S4: Compare the actual power load data with the power load forecast data output by the load forecast model, and dynamically adjust the parameters of the local forecast model based on the comparison results to optimize the accuracy of the power load forecast data and power allocation strategy.

2. The urban power distribution method based on load forecasting according to claim 1, characterized in that: The process of collecting power load data from various areas of the city through smart meters is as follows: Electricity load data from various areas of the city is collected using smart meters to form a raw electricity load data sequence. ,in, This represents the load value at the nth time point.

3. The urban power distribution method based on load forecasting according to claim 1, characterized in that: The specific process for calculating the sequence gradient is as follows: Calculate the time series gradient of the power load data to capture the rate of load change. For each time point t of the power load data, calculate the sequence gradient. Defined as: ,in, The gradient represents the magnitude of load change at adjacent time points. The power load data value at time point t. For the time point is The larger the absolute value of the sequence gradient of the power load data, the more severe the load fluctuation.

4. The urban power distribution method based on load forecasting according to claim 1, characterized in that: The specific process of dynamically adjusting the noise injection amplitude based on the sequence gradient is as follows: Based on sequence gradient The noise injection amplitude is dynamically adjusted, specifically during periods of high change rate. Injecting less noise preserves more data details, especially during periods of low change rate. Injecting significant noise enhances privacy protection; noise amplitude The adjustment formula is: ,in, For the initial privacy budget, The adjustable parameters are used to control the sensitivity of noise to changes with the gradient; noise amplitude. The adjustment formula uses the sigmoid function to achieve a nonlinear mapping between the sequence gradient and noise, ensuring that high sequence gradients are avoided. near When the sequence gradient is low Approaching 0.

5. The urban power distribution method based on load forecasting according to claim 1, characterized in that: The specific process of injecting Laplace noise into the power load data value to generate desensitized power load data is as follows: For each power load data value Injecting Laplace noise Generate desensitized data : Among them, Laplace noise Follows a Laplace distribution: ,in, The sensitivity parameter is usually set to 1, which represents the maximum magnitude of data variation. This is a dynamically adjusted privacy budget.

6. The urban power distribution method based on load forecasting according to claim 1, characterized in that: The specific process by which the edge nodes collect and aggregate the de-identified power load data from each region is as follows: Edge nodes collect anonymized power load data from each region. ,in, This represents the desensitized load value at the nth time point; the desensitized power load data collected from each region by the edge nodes are summed to form a regional load dataset.

7. The urban power distribution method based on load forecasting according to claim 1, characterized in that: The specific process of training a local load prediction model using a fusion model of federated learning and knowledge distillation, first defining the global model and the edge model, is as follows: Define the global model as: denoted as It is a high-performance, complex model maintained in the cloud, possessing knowledge of global data training; Define the edge model as: denoted as It is a lightweight model deployed locally on edge nodes.

8. The urban power distribution method based on load forecasting according to claim 1, characterized in that: The specific process by which the cloud receives the local prediction model parameters trained by each edge node and summarizes and updates the global model is as follows: The cloud receives the local prediction model parameters trained by each edge node. Where m is the number of edge nodes, the global model parameters are updated through a federated learning aggregation strategy. The specific process is as follows: Weighted aggregation of each edge model based on regional load scale: ,in, Let i be the amount of load data covered by the i-th edge node. This represents the total global data volume. These are the parameters for the i-th local prediction model.

9. The urban power distribution method based on load forecasting according to claim 1, characterized in that: The multi-source data specifically includes: Meteorological data, economic indicators, and historical power load data; among which, meteorological data includes: temperature, humidity, and sunshine; economic indicators include: GDP growth rate and industrial output value.

10. A city power distribution system based on load forecasting, characterized in that: include: Privacy-aware data preprocessing module: Collects power load data from various areas of the city through smart meters, performs sequence gradient calculation on the collected power load data, dynamically adjusts the noise injection amplitude according to the sequence gradient, and injects Laplace noise into the power load data values ​​to generate desensitized power load data; Edge Federated Knowledge Distillation Training Module: Edge nodes collect and aggregate desensitized power load data from various regions, and use a federated learning and knowledge distillation fusion model to train the local load prediction model. First, a global model and an edge model are defined. Then, the knowledge of the global model is transferred to the local prediction model through federated knowledge distillation, which improves the performance of the local prediction model without sharing the original power load data. The cloud-based interpretable prediction decision module receives the local prediction model parameters trained by each edge node, summarizes and updates the global model, combines multi-source data, uses the interpretable-enhanced load prediction model to output power load prediction data, and generates power allocation strategies based on the power load prediction data. Closed-loop feedback optimization module: Compares the actual power load data with the power load forecast data output by the load forecast model, and dynamically adjusts the parameters of the local forecast model based on the comparison results to optimize the accuracy of the power load forecast data and power allocation strategy.

Images