Intelligent load management and allocation method for electric meter box

Abstract

The invention discloses an intelligent load management and allocation method for an electric meter box, relates to the technical field of intelligent power grid management and energy optimization, and solves the problems that the adaptability of load management in an existing electric meter box is insufficient, the prediction accuracy is influenced by data prejudice, and the dynamic adjustment capability is limited. According to the method, diversified data are collected and analyzed through a real-time data interface and a gradient boosting tree technology, a power consumption prediction model is constructed in combination with a long-short term memory (LSTM) network, and meanwhile, prediction performance is optimized by adopting a data enhancement algorithm and an adaptive enhancement algorithm; fast response and dynamic adjustment are realized by using a sliding window prediction method based on Kalman filtering, a load distribution strategy is optimized through an adaptive depth deterministic strategy gradient algorithm, and collaborative distribution scheduling and management are performed on cross-regional electric meter boxes through a distributed collaborative scheduling system; according to the invention, the stability and efficiency of the power system are greatly improved, and the energy utilization efficiency and economic benefits are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of smart grid management and energy optimization, and more specifically to a method for smart load management and allocation of an electric meter box. Background Art

[0002] With the development of smart grid technology and the popularization of Internet of Things technology, the intelligence of meter boxes has become an important part of the modernization of power systems. In the current era, meter boxes not only have to bear the traditional function of electricity metering, but also need to integrate more intelligent functions, such as remote monitoring, fault warning, data collection and analysis, etc., to meet the power system's needs for high efficiency, high reliability and high safety.

[0003] In the prior art, smart meter boxes realize online active monitoring of the working condition, safety and environmental status of power meter boxes by integrating multiple sensors and communication technologies. For example, in the article "Introducing Edge Intelligence for Smart Meters through Federated Segmentation Learning" published by Professor Wang Yi, an innovative cloud-edge-end architecture is proposed by combining federated learning and segmentation learning, which effectively solves the dual limitations of smart meters in hardware resources and data resources. In addition, in patent CN118487385B, the output end switch is controlled by switching when the power at the access end is abnormal, and each abnormal load branch is confirmed and disconnected. At the same time, the abnormal power loss branches at the output end are confirmed by combining the power consumption relationship between the output end and each load branch; however, although the prior art has made certain progress in the load management and allocation of smart meter boxes, it still has the following shortcomings:

[0004] First, the power system is in a highly dynamic environment, and factors such as weather, seasons, and holidays can significantly affect electricity consumption patterns. Traditional forecasting models often assume that the environment is relatively stable, but in reality these factors change very frequently and unpredictably. For example, extreme weather events (such as heat waves or cold waves) may cause a sudden surge in electricity consumption, and the model fails to adjust the forecast results in time, resulting in load management failure. Second, forecasting models usually rely on historical data for training, but these data may be biased. For example, if the training data mainly comes from a specific region or time period, the model may not be able to accurately predict the load conditions in other regions or time periods. This bias will cause the model to perform poorly when faced with new situations; finally, forecasting models are usually updated based on fixed time windows (such as every hour or every day), while in actual operation, load changes may be very rapid. If the feedback mechanism is not timely, the system may not be able to respond to these changes quickly, resulting in the failure of the control strategy. Summary of the invention

[0005] In view of the deficiencies of the prior art, the present invention discloses an intelligent load management and allocation method for an electric meter box to solve the problems raised in the background technology.

[0006] The present invention adopts the following technical solutions:

[0007] A method for intelligent load management and deployment of an electric meter box, comprising the following steps:

[0008] S1. Collect and analyze diversified data through real-time data interface and gradient boosting tree, including real-time load data, historical power consumption data, weather forecast, seasonal information, holiday schedule, to form a comprehensive data set; the historical power consumption data includes power consumption in different time periods and under different weather conditions;

[0009] S2. Based on the comprehensive data set, a power consumption prediction model is constructed through the long short-term memory network LSTM to predict power consumption in real time;

[0010] S3, based on the historical electricity consumption data in S1, enhances the historical electricity consumption data by synthesizing minority class oversampling, eliminates regional and time period bias, and trains a strong prediction model through an adaptive enhancement algorithm;

[0011] S4, based on the strong prediction model output by S3 and the real-time load data output by S1, sets the initial time window through the sliding window prediction method based on Kalman filtering, dynamically adjusts the window size according to the real-time rate of load change, continuously predicts the power demand, and updates the prediction results in real time according to the new observation value;

[0012] S5, based on the prediction results of S4 and the real-time load data output by S1, dynamically adjust the load distribution strategy through the adaptive deep deterministic policy gradient algorithm;

[0013] S6. Monitor load changes in real time. If the load change exceeds a preset threshold, load allocation is performed through an instant allocation algorithm.

[0014] S7. Coordinated allocation, scheduling and management of cross-regional electricity meter boxes are performed through a distributed collaborative scheduling system, wherein the distributed collaborative scheduling system performs collaborative scheduling of cross-regional electricity meter boxes through an alternating direction multiplier method.

[0015] As a further technical solution of the present invention, the analysis method of analyzing diversified data by gradient boosting tree in S1 is as follows: first, the real-time load data of the meter box, the historical power consumption data and the diversified information of the external data source are integrated through the real-time data interface, and the integrated data is preprocessed by outlier detection and normalization processing; the external data source includes weather forecast, seasonal information, and holiday schedule; based on the preprocessed data set, a series of basic decision trees are generated as weak learners, and each weak learner independently performs preliminary fitting on the data; in the iterative process, the residual of the previous round of prediction results is calculated by the gradient descent method, and the training set is updated according to the residual. training target; if the residual is greater than the preset threshold, the weak learner is regenerated; in the new round of iteration, the nodes of the decision tree are split according to the importance of the features through feature weighting and splitting strategies, and the weights of the nodes are adjusted; after the iteration meets the preset threshold, based on the iteratively generated weak learners, the weak learners generated in each round of iteration are integrated into the strong learner through the ensemble learning method to form a gradient boosting tree model; then the loss value of the current gradient boosting tree model is calculated through the mean square error MSE, and the parameters of the gradient boosting tree model are adjusted according to the loss value; finally, the gradient boosting tree model is regularized through L1 regularization, and the decision tree is pruned.

[0016] As a further technical solution of the present invention, the power consumption prediction model includes an input layer, an embedding layer, an LSTM layer, an attention mechanism layer, a fully connected layer and an output layer; the working method of the power consumption prediction model is:

[0017] F1. receiving multi-source data in a comprehensive data set through the input layer, and the input layer ensures the format and quality of the data through normalization and standardization methods;

[0018] F2, converting discrete category data into continuous vector representations through the embedding layer; the embedding layer maps each category to a fixed-length vector through a lookup table to capture the characteristics of the category data, the category data including holiday schedules and seasonal information;

[0019] F3, capturing long-term dependencies in time series data through the LSTM layer, which controls the inflow and outflow of information through input gates, forget gates, and output gates, dynamically adjusts the degree of memory of historical information, and captures long-term dependencies in time series data through cell states;

[0020] F4, highlighting important time steps and features through the attention mechanism layer; the attention mechanism layer generates an attention map by calculating the weight of each time step and feature;

[0021] F5. The outputs of the LSTM layer and the attention mechanism layer are integrated through the fully connected layer to generate a high-dimensional feature vector;

[0022] F6. Generate a power consumption prediction result through the output layer; the output layer performs a nonlinear transformation on the high-dimensional feature vector through the activation function ReLU, and converts the prediction result into an actual power consumption value through a normalization operation.

[0023] As a further technical solution of the present invention, the method for enhancing the historical electricity consumption data by synthetic minority oversampling in S3 is as follows: first, the historical electricity consumption data is preliminarily divided by a density-based clustering algorithm to identify electricity consumption patterns in different regions and time periods, including minority samples and all samples; the minority samples represent data points in time periods and regions with low or high electricity consumption, and based on the identified minority samples, the Euclidean distance between the minority samples and all other samples is calculated by a nearest neighbor-based interpolation method to find the K nearest neighbor samples of the minority samples; according to the distribution of the K nearest neighbor samples, one of them is selected for interpolation to generate a new minority sample.

[0024] As a further technical solution of the present invention, the working method of the adaptive enhancement algorithm is as follows: based on the enhanced data set, the adaptive enhancement algorithm first assigns weights to each sample in the data set through an initial weight assignment method, then trains the first weak prediction model based on the current weight distribution, and calculates the prediction error of the weak prediction model on the training set; according to the prediction error of the weak prediction model, calculates the error rate of each sample; according to the error rate, dynamically adjusts the weight of each sample through a weight adjustment strategy, focuses on fuzzy samples, and performs multiple iterative training; during the iterative process, the adaptive enhancement algorithm dynamically adjusts the weight of the next round of weak prediction models according to the prediction error of each weak prediction model on the training set, and weighted combines the adjusted models in each round to generate a strong prediction model; during each iterative training process, the adaptive enhancement algorithm also dynamically adjusts the weight of the input feature according to the feature importance score of the weak prediction model through a feature importance evaluation mechanism.

[0025] As a further technical solution of the present invention, the working method of the sliding window prediction method based on Kalman filtering is: by reading the initial values ​​of historical power consumption data and real-time load data, setting the initial size of the sliding window, initializing the state vector x0 and covariance matrix p0 of the Kalman filter, wherein x0 represents the power demand forecast value at the initial moment, and p0 represents the uncertainty of the initial state; if the current moment is t, the state vector x0 at the previous moment is used to predict the power demand at the previous moment through the prediction equation of the Kalman filter. t-1 And the process noise covariance matrix Q predicts the current state vector x t And the covariance matrix p at the current moment t; When the real-time rate of load change is higher than or equal to the preset threshold, the sliding window size is reduced through the dynamic window adjustment mechanism; when the real-time rate of load change is lower than the preset threshold, the sliding window size is expanded; at each time step, the Kalman filter-based sliding window prediction method updates the prediction result according to the current observation value and the state of the prediction model through the Kalman filter algorithm; and after each observation update, the new data is integrated into the sliding window through the data fusion in the sliding window, and the oldest data in the window is removed; if the current moment has not reached the end time of the prediction period, the above method is repeated to achieve continuous prediction.

[0026] As a further technical solution of the present invention, the working method of the adaptive deep deterministic policy gradient algorithm is:

[0027] T1. Construct an adaptive actor network and an adaptive critic network; the adaptive actor network is used to generate a load distribution strategy according to the current load state of the meter box, and the adaptive critic network is used to predict the reward obtained after adopting the load distribution strategy;

[0028] T2, initializing the adaptive actor network, the adaptive critic network and the experience pool, wherein the experience pool is used to store historical experience;

[0029] T3, for each time step t, receiving the real-time load data l output by S1 t And the prediction result p output by S4 t , combined with the current load state s t , through the deep convolutional neural network to the current load state s t Perform feature extraction to obtain high-dimensional state representation

[0030] T4, the adaptive actor network is characterized according to the current state and the current policy parameter θ art , generate load distribution strategy α t ;In the process of generating load distribution strategy, the size of exploration noise is adjusted through adaptive exploration mechanism to balance the relationship between exploration and utilization;

[0031] T5, generated load distribution strategy α t Adjust load distribution through wireless communication protocol executed on the meter box;

[0032] T6. After executing the action, the adaptive deep deterministic policy gradient algorithm receives the reward r t+1 and the next state s t+1 , the tuple (s t ,α t ,r t+1 ,s t+1) is stored in the experience pool;

[0033] T7. During the training process, experience is sampled from the experience pool through an adaptive sampling strategy for training. The adaptive sampling strategy is weighted according to the timestamp and the reward value to improve the sample utilization efficiency;

[0034] T8. Based on the sampled experience, the adaptive critic network adjusts the value function parameters θ according to the current value function parameters θ. cri Calculate the target value and update the value function parameter θ based on the gradient descent algorithm cri , the value function parameter θ cri During the update process, the adaptive critic network calculates the gradient of the actor network through the policy gradient function. The adaptive actor network updates the actor network gradient information provided by the adaptive critic network through the adaptive learning rate δ art Update the policy parameters θ art ; The adaptive learning rate is dynamically adjusted according to the historical learning efficiency and the current strategy stability; wherein the formula expression of the strategy gradient function is:

[0035]

[0036] In formula (1), Represents the objective function J of the actor network to the current policy parameter θ art The gradient of E is used to update the parameters of the actor network. dt Indicates the current load status s t The expected value in the experience pool, which stores historical interaction data, including the four-tuple of state, action, reward and next state (s t ,α t ,r t+1 ,s t+1 ); Represents the gradient of the adaptive critic network to action a, which is used to reflect the current state s t Next, the impact of action a on the Q value; Indicates that in state s t The action generated by the actor network U is: art The gradient of is used to reflect the influence of the parameters of the policy network on the generated actions;

[0037] T9. At the end of each training cycle, the parameters of the target network are updated according to the current network parameters and the soft update coefficient τ through the target network update mechanism; the soft update coefficient τ is dynamically adjusted according to the fluctuations during the training process to balance the difference between the current network and the target network;

[0038] T10, iterative loop T3 to T9, terminated when the preset training termination conditions are reached; the training termination conditions include the number of training rounds and strategy convergence; during the training process, the early stopping mechanism is used to avoid overfitting and insufficient training.

[0039] As a further technical solution of the present invention, the working method of the real-time allocation algorithm is: to determine whether the load change exceeds the preset threshold through a threshold comparison mechanism, if not, skip the execution step of the real-time allocation algorithm; if exceeded, generate a priority list through a load distribution priority mechanism, and calculate the load adjustment amount through a load balancing algorithm based on the priority list. The load balancing algorithm determines the optimal adjustment amount of each load device through a linear programming model, and then sends the load adjustment amount to the load device through the control protocol Modbus to adjust the power or switch state of the load device.

[0040] As a further technical solution of the present invention, the distributed collaborative scheduling system includes a data access module, an optimization scheduling module and a communication coordination module; the data access module obtains the load data, environmental data and equipment status data of the meter boxes in each area through the data interface, and fills missing values, processes outliers and normalizes the data through the data preprocessing method; the optimization and scheduling module performs collaborative scheduling of cross-regional meter boxes through the alternating direction multiplier method, and outputs the scheduling strategy to the communication coordination module; the communication coordination module includes a distributed communication unit and a state synchronization unit; the distributed communication unit realizes efficient communication between the meter boxes in each area through a message queue; the state synchronization unit ensures the synchronization of the status information of the meter boxes in each area through heartbeat detection, and performs redundant backup and fault switching through a fault-tolerant mechanism.

[0041] As a further technical solution of the present invention, the working method of the alternating direction multiplier method is:

[0042] R1. Define the global objective function Among them, r i represents the load distribution of the electric meter box in the i-th area, f(r i ) represents the energy consumption cost and load balance degree of the i-th area; N represents the upper limit of the area; the global objective function includes multiple sub-objective functions, each sub-objective function corresponds to an electric meter box in an area;

[0043] R2. Decompose the global optimization problem into multiple sub-problems through auxiliary variables z and Lagrange multiplier γ. Each sub-problem corresponds to the meter box in a region. The decomposition formula is:

[0044]

[0045] In formula (2), z represents the auxiliary variable, which is used to ensure that the load distribution of all regions is consistent at the global level; γ represents the Lagrange multiplier, which is used to deal with global constraints; β represents the penalty parameter, which is used to balance the weight of local optimization and global constraints;

[0046] R3. In each iteration step, the load distribution of the meter box in each area is first updated by the local update function r i , ensuring that the load distribution of the electric meter boxes in each area is optimized in the current iteration, the formula expression of the local update function is:

[0047]

[0048] In formula (3), l represents the number of iterations; Represents the dual variable, which is used to transmit information during the iteration process;

[0049] R4. Update the auxiliary variable z through the global update function to ensure that the load distribution of the meter boxes in all regions is consistent at the global level; the formula expression of the global update function is:

[0050]

[0051] In formula (4), z k+1 Represents the updated auxiliary variables; Indicates the updated load distribution of the electric meter boxes in each area;

[0052] R5. Update the Lagrange multiplier γ through the dual update function to ensure that the load distribution of the meter boxes in all regions is consistent at the global level. The formula expression of the dual update function is:

[0053]

[0054] In formula (5), γ k+1 represents the updated Lagrange multiplier;

[0055] R6. During the iterative update process, if the change in the solution of the sub-problem is greater than the preset threshold, the penalty parameter β is increased; if in a certain iteration, if the change in the solution of the sub-problem is less than the preset threshold, the penalty parameter β is reduced.

[0056] Based on the above technical solution, the positive and beneficial effects of the present invention are:

[0057] 1. The present invention comprehensively collects and analyzes diversified data including real-time load data, historical power consumption data, weather forecasts, seasonal information, and holiday arrangements through real-time data interfaces and gradient boosting tree technology. This comprehensive data input method enables the prediction model to more accurately capture the dynamic changes in the power system, especially the impact of unpredictable factors such as extreme weather events on power consumption patterns, effectively solving the problem of insufficient adaptability of the prediction model to dynamic environments and improving the accuracy and effectiveness of load management.

[0058] 2. The present invention uses synthetic minority class oversampling technology for data enhancement processing, which effectively eliminates data bias; at the same time, through the application of adaptive enhancement algorithm, the generalization ability of the model is further improved, so that it can still maintain a high prediction accuracy when facing unknown or complex situations. This strategy not only improves the prediction performance of the model, but also enhances the overall effectiveness of load management.

[0059] 3. The present invention adopts a sliding window prediction method based on Kalman filtering, which can dynamically adjust the window size according to the real-time rate of load change to achieve continuous prediction and real-time update; this dynamic adjustment mechanism enables the system to quickly capture the trend of load changes and adopt corresponding control strategies for adjustment. Therefore, the present invention effectively solves the problem that the traditional prediction model cannot respond to rapid load changes in time due to fixed time window updates, and improves the response speed and stability of the system.

[0060] 4. The present invention dynamically adjusts the load distribution strategy through an adaptive deep deterministic policy gradient algorithm, thereby achieving the rational allocation and utilization of energy resources; this optimization strategy not only reduces energy waste, improves energy efficiency, but also reduces the operating cost of the power system. At the same time, through real-time monitoring of load changes and the application of instant deployment algorithms, the flexibility and reliability of the system are further improved, providing strong support for the sustainable development of the power industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A schematic diagram of the steps of an electric meter box intelligent load management and deployment method of the present invention;

[0062] Figure 2 A working method step diagram of the power consumption prediction model of the present invention;

[0063] Figure 3 This is a schematic diagram of the principle framework of the gradient boosting tree for analyzing diversified data in the present invention;

[0064] Figure 4 It is a working method framework diagram of the sliding window prediction method based on Kalman filtering of the present invention;

[0065] Figure 5This is a framework diagram of the working method of the real-time deployment algorithm of the present invention. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] In the present invention, if Figure 1 As shown: The steps of a method for intelligent load management and allocation of an electric meter box are as follows:

[0068] S1. Collect and analyze diversified data through real-time data interface and gradient boosting tree, including real-time load data, historical power consumption data, weather forecast, seasonal information, holiday schedule, and form a comprehensive data set; the historical power consumption data includes power consumption in different time periods and under different weather conditions; among them, Figure 3 As shown in FIG. 1 , the analysis method of analyzing diversified data by gradient boosting tree in S1 is as follows: first, the real-time load data of the meter box, the historical power consumption data and the diversified information of the external data source are integrated through the real-time data interface, and the integrated data is preprocessed by outlier detection and normalization processing; the external data source includes weather forecast, seasonal information, and holiday schedule; based on the preprocessed data set, a series of basic decision trees are generated as weak learners, and each weak learner independently performs preliminary fitting on the data; in the iterative process, the residual of the previous round of prediction results is calculated by the gradient descent method, and the training target is updated according to the residual; if the residual If it is greater than the preset threshold, the weak learner is regenerated; in the new round of iteration, the nodes of the decision tree are split according to the importance of the features through feature weighting and splitting strategy, and the weights of the nodes are adjusted; after the iteration meets the preset threshold, based on the iteratively generated weak learners, the weak learners generated in each round of iteration are integrated into the strong learner through the ensemble learning method to form a gradient boosting tree model; then the loss value of the current gradient boosting tree model is calculated by the mean square error MSE, and the parameters of the gradient boosting tree model are adjusted according to the loss value; finally, the gradient boosting tree model is regularized by L1 regularization, and the decision tree is pruned.

[0069] In a specific embodiment, real-time load data is collected in real time through sensors or communication modules built into the meter box. These data reflect the current load status of the power grid and are the basis for load management and deployment. Historical electricity consumption data is extracted from the power company's database, including electricity consumption in different time periods and weather conditions in the past, which is used to analyze electricity consumption trends and patterns. Weather forecasts are obtained through meteorological service APIs or partners, including meteorological information such as temperature, humidity, and wind speed. This information has a direct impact on electricity demand. For example, high temperatures will increase air conditioning electricity consumption. Seasonal information is automatically generated according to the calendar or preset rules to reflect the current season. Electricity consumption habits and demands vary in different seasons. Holiday schedules are obtained from government or public databases. Holidays usually affect electricity consumption patterns. For example, peak electricity consumption on weekdays may be transferred to rest days.

[0070] The core of the gradient boosting tree is to integrate multiple basic decision trees (weak learners) to build a strong learner. Each basic decision tree fits the data independently, and then their prediction results are combined by weighted voting or averaging to improve the overall prediction accuracy. This method can make full use of the advantages of different weak learners and reduce the bias and variance of a single model. When building the model, first generate multiple training subsets by random sampling or feature sampling, and then train the basic decision trees on these subsets respectively. In the prediction stage, each basic decision tree will give a prediction result, and finally these results are combined by an ensemble learning algorithm to obtain the final prediction value. This method can significantly improve the prediction performance of the model while reducing the risk of overfitting.

[0071] During the iteration process, the gradient boosting tree uses the gradient descent method to optimize the parameters of the model. Specifically, it calculates the residual (i.e., error) between the current prediction result and the actual result, and then updates the parameters of the model based on the size and direction of the residual to reduce the error. In each iteration, the gradient boosting tree generates a new basic decision tree based on the prediction results and residuals of the previous round and adds it to the model. This new basic decision tree will try to correct the errors in the previous round of predictions, thereby improving the overall prediction accuracy. Through multiple iterations, the model can gradually approach the actual electricity demand and improve the accuracy of the prediction.

[0072] When constructing the basic decision tree, the gradient boosting tree splits the nodes according to the importance of the features. Specifically, it selects those features that have the greatest impact on the prediction results as splitting conditions to generate a more accurate prediction model. In the feature selection stage, the gradient boosting tree calculates indicators such as information gain or Gini impurity for each feature to evaluate its importance. Then, when constructing the basic decision tree, it prioritizes those features with higher importance as splitting conditions. In this way, the model can more accurately capture the relationship between electricity demand and features, improving the accuracy and robustness of the prediction.

[0073] Compared with the existing technology, the gradient boosting tree can significantly improve the prediction accuracy of the model by integrating multiple basic decision trees to build a strong learner. In addition, through strategies such as regularization and feature weighting, the gradient boosting tree can reduce the risk of overfitting of the model and improve the generalization ability of the model. Secondly, the gradient boosting tree can build a personalized prediction model based on different data sets and features to adapt to the electricity demand in different scenarios. At the same time, through strategies such as gradient descent and feature weighting, the gradient boosting tree can continuously optimize the parameters and structure of the model during the iteration process, improving the performance and stability of the model.

[0074] S2. Based on the comprehensive data set, a power consumption prediction model is constructed through a long short-term memory network LSTM to predict power consumption in real time; wherein the power consumption prediction model includes an input layer, an embedding layer, an LSTM layer, an attention mechanism layer, a fully connected layer and an output layer; the working method of the power consumption prediction model is as follows: Figure 2 As shown: F1, receiving multi-source data in the comprehensive data set through the input layer, the input layer ensures the format and quality of the data through normalization and standardization methods; F2, converting discrete category data into continuous vector representation through the embedding layer; the embedding layer maps each category to a fixed-length vector through a lookup table to capture the characteristics of the category data, and the category data includes holiday arrangements and seasonal information; F3, capturing long-term dependencies in time series data through the LSTM layer, the LSTM layer controls the inflow and outflow of information through input gates, forget gates and output gates, dynamically adjusts the degree of memory of historical information, and captures long-term dependencies of time series data through cell states; F4, highlighting important time steps and features through the attention mechanism layer; the attention mechanism layer generates an attention map by calculating the weight of each time step and feature; F5, synthesizing the outputs of the LSTM layer and the attention mechanism layer through the fully connected layer to generate a high-dimensional feature vector; F6, generating power consumption prediction results through the output layer; the output layer performs nonlinear transformation on the high-dimensional feature vector through the activation function ReLU, and converts the prediction results into actual power consumption values ​​through normalization operations.

[0075] In a specific embodiment, the embedding layer is based on distributed representation, that is, by mapping each category to a vector in a high-dimensional space, similar categories are closer in the vector space, and different categories are farther away. This representation is more efficient and expressive than the traditional one-hot encoding because it can capture the potential relationship between categories. In the electricity consumption prediction model, the embedding layer converts category data such as holidays and seasons into continuous vectors, providing richer input features for the subsequent LSTM layer. The core of LSTM lies in its cell state and three gating mechanisms (input gate, forget gate, output gate). The cell state is responsible for the storage and transmission of long-term information, while the three gating mechanisms control the input of new information, the forgetting of old information, and the selection of output information respectively. This design enables LSTM to capture long-term dependencies in time series data, which is very suitable for data prediction with obvious periodicity and seasonal changes such as electricity consumption. In the electricity consumption prediction model, the attention mechanism generates an attention map by calculating the importance weight of each time step or feature, thereby highlighting the time steps and features that contribute significantly to the prediction results. This mechanism enables the model to process input data more flexibly and improve the accuracy and robustness of predictions.

[0076] Compared with the existing technology, the electricity consumption prediction model can efficiently process multi-source, heterogeneous input data, including categorical data and numerical data, through the combination of embedding layer and LSTM layer in application implementation, which improves the accuracy and robustness of prediction. Secondly, the design of LSTM layer enables the model to capture long-term dependencies in time series data, which is particularly important for the prediction of data with obvious periodicity and seasonal changes such as electricity consumption. In addition, the introduction of attention mechanism enables the model to process input data more flexibly, highlighting the time steps and features that have important contributions to the prediction results, and improving the accuracy and generalization ability of prediction.

[0077] S3. Based on the historical electricity consumption data in S1, the historical electricity consumption data is enhanced by synthetic minority oversampling to eliminate regional and time period bias, and a strong prediction model is trained by an adaptive enhancement algorithm. The method for enhancing the historical electricity consumption data by synthetic minority oversampling in S3 is as follows: first, the historical electricity consumption data is preliminarily divided by a density-based clustering algorithm to identify electricity consumption patterns in different regions and time periods, including minority samples and all samples; the minority samples represent data points in time periods and regions with low or high electricity consumption. Based on the identified minority samples, the Euclidean distance between the minority samples and all other samples is calculated by a nearest neighbor-based interpolation method to find the K nearest neighbor samples of the minority samples; according to the distribution of the K nearest neighbor samples, one of them is selected for interpolation to generate a new minority sample.

[0078] In a specific embodiment, synthetic minority oversampling SMOTE is a technology widely used in processing unbalanced data sets. Its main purpose is to increase the number of minority samples by synthesizing new samples, so that the model can learn the characteristics of the minority class more fully during the training process. The core idea is to find the nearest neighbor of the minority sample in the feature space, and then randomly select a point along the line between these samples as the position of the new sample. This not only increases the number of minority samples, but also maintains the spatial distribution characteristics of the original samples, avoiding the risk of overfitting caused by simple replication.

[0079] Density-based clustering algorithm DBSCAN is a density-based spatial clustering algorithm that can discover clusters of arbitrary shapes. DBSCAN distinguishes different data points by defining "core objects", "boundary objects" and "noise points". A point is considered a core object if the number of points contained in its neighborhood exceeds a certain threshold; if a point is not a core object but belongs to the neighborhood of a core object, it is considered a boundary object; a point that is neither a core object nor in the neighborhood of any core object is considered a noise point. DBSCAN does not require the number of clusters to be specified in advance and can handle noisy data sets well.

[0080] In practical applications, SMOTE technology can identify data points in time periods and regions with low or high electricity consumption, and generate new minority class samples through interpolation methods, thereby balancing the number of samples of different categories in the data set. This data enhancement processing not only eliminates regional and time bias, but also enables the model to learn the characteristics of different electricity consumption patterns more comprehensively during training.

[0081] Compared with the existing technology, SMOTE technology can automatically identify and enhance minority class samples, thereby balancing the number of samples of different categories in the data set and solving the problem of regional and time period bias. Secondly, by generating new minority class samples through interpolation methods, it not only increases the diversity of data, but also improves the generalization ability and prediction accuracy of the model. In addition, the combination of SMOTE technology and adaptive enhancement algorithm enables the model to learn the characteristics of different power consumption patterns more comprehensively during the training process, further improving the prediction performance. Finally, this technology has high flexibility and scalability, and can adjust parameters and expand functions according to actual needs to meet the application needs in different scenarios. In summary, the application of SMOTE technology in the intelligent load management and dispatching method of meter boxes has significant advantages and characteristics, and provides strong technical support for decisions such as power grid dispatching and energy distribution.

[0082] Furthermore, the working method of the adaptive enhancement algorithm is as follows: based on the enhanced data set, the adaptive enhancement algorithm first assigns weights to each sample in the data set through an initial weight assignment method, and then trains the first weak prediction model based on the current weight distribution, and calculates the prediction error of the weak prediction model on the training set; according to the prediction error of the weak prediction model, the error rate of each sample is calculated; according to the error rate, the weight of each sample is dynamically adjusted through a weight adjustment strategy, focusing on fuzzy samples, and performing multiple iterative training; during the iterative process, the adaptive enhancement algorithm dynamically adjusts the weight of the next round of weak prediction models according to the prediction error of each weak prediction model on the training set, and weightedly combines the adjusted models in each round to generate a strong prediction model; during each iterative training process, the adaptive enhancement algorithm also dynamically adjusts the weight of the input feature according to the feature importance score of the weak prediction model through a feature importance evaluation mechanism.

[0083] In a specific embodiment, the initial weight assignment method refers to assigning an initial weight to each sample in the data set at the beginning of the adaptive boosting algorithm. The initial weights are the same to ensure that each sample contributes equally to the model training at the beginning of the algorithm. Weak prediction models refer to those models with weak prediction capabilities when used alone. In the adaptive boosting algorithm, weak prediction models are usually trained by simple machine learning algorithms (such as decision trees). Each weak prediction model is trained according to the current sample weight distribution, focusing on capturing local features in the data. The prediction error refers to the difference between the prediction result of the weak prediction model on the training set and the true label. The error rate refers to the proportion of samples with incorrect predictions to the total samples. These two indicators are used to evaluate the performance of the weak prediction model and guide subsequent weight adjustments. The weight adjustment strategy adjusts the sample weights so that the model pays more attention to samples that made mistakes in the previous round of predictions in subsequent iterations. Specifically, the weight of samples with incorrect predictions will increase, while the weight of samples with correct predictions will decrease. This can prompt the model to pay more attention to samples that are difficult to predict in subsequent iterations, thereby gradually improving the overall prediction performance. The feature importance evaluation mechanism is used to evaluate the contribution of each feature in the prediction. During each iterative training process, the adaptive boosting algorithm dynamically adjusts the weights of the input features based on the feature importance scores of the weak prediction model. This helps the model pay more attention to features that have a greater impact on the prediction results, thereby improving the model's prediction accuracy.

[0084] In practical applications, the AdaBoost algorithm improves the prediction accuracy and robustness of the model by enhancing the fuzzy samples in the historical electricity consumption data. Specifically, the algorithm can identify samples that are misclassified by multiple weak prediction models (i.e., fuzzy samples), and dynamically adjust the sample weights so that these samples receive more attention in the subsequent training process. This mechanism helps the model learn more complex electricity consumption patterns, thereby improving the accuracy of the prediction. At the same time, through the feature importance evaluation mechanism, the algorithm can identify features that have important contributions to the prediction results and dynamically adjust their weights to further optimize the performance of the model. Therefore, in the intelligent load management and deployment method of the meter box, the application of the AdaBoost algorithm can significantly improve the prediction performance of the model and provide a scientific basis for decisions such as power grid scheduling and energy distribution.

[0085] Compared with the existing technology, the application of AdaBoost algorithm in the intelligent load management and dispatching method of meter box shows the following advantages and characteristics: First, the algorithm can automatically identify and enhance fuzzy samples, that is, samples that are misclassified by multiple weak prediction models, thereby improving the prediction accuracy and robustness of the model. Secondly, by dynamically adjusting the sample weights and feature weights, the algorithm can learn more complex power consumption patterns and further optimize the performance of the model. In addition, as an integrated learning method, the AdaBoost algorithm has high flexibility and scalability, and can adjust parameters and expand functions according to actual needs. Finally, the algorithm can gradually reduce the prediction error and improve the generalization ability of the model during the training process, thereby showing better prediction performance in practical applications.

[0086] In the specific implementation, the hardware working environment configuration of the adaptive enhancement algorithm is as follows: server: CPU: Intel Xeon E5-2680 v4 (14 cores, 2.40GHz); memory: 64GB DDR4; storage: 1TB SSD; operating system: Ubuntu 20.04 LTS; smart meter: support Modbus protocol, collect electricity consumption data every 15 minutes; sensor: temperature, humidity, light and other environmental sensors, collect data once an hour; switch: Gigabit Ethernet switch; router: enterprise-level router, support IPv6; laptop: used for monitoring and management, equipped with Intel Core i7 processor, 16GB RAM, 512GB SSD; mobile phone and tablet: used for remote access and control, support Android and iOS systems.

[0087] Based on the above hardware environment, a comparative experiment was designed to verify the effect of the adaptive enhancement algorithm in the intelligent load management and deployment method of the meter box. The experiment was divided into two groups: Group A used the adaptive enhancement algorithm, and Group B used the traditional linear regression algorithm. Both groups of experiments were conducted under the same hardware and experimental conditions, with five experiments in each group. The main goal of the experiment is to compare the performance of the two groups of algorithms in terms of accuracy and stability of electricity consumption prediction. The data is the historical electricity consumption data of a certain area in the past year, including environmental data such as electricity consumption, temperature, humidity, and light every 15 minutes. The data set is divided into a training set (80%) and a test set (20%). The experimental record table is shown in Table 1:

[0088] Table 1 Adaptive enhancement algorithm comparison experiment record

[0089]

[0090]

[0091] As can be seen from the data table 1, the RMSE and MAE of group A are 0.32kWh and 0.21kWh respectively, while the RMSE and MAE of group B are 0.55kWh and 0.38kWh respectively. This shows that the adaptive boosting algorithm can more accurately capture the changing trend of electricity consumption, especially when dealing with complex, nonlinear data. Although the running time of the adaptive boosting algorithm is slightly longer than that of the linear regression algorithm, its advantage in prediction accuracy far outweighs this. The average running time of group A is 120 seconds, while the average running time of group B is 82 seconds. Considering the high requirements for prediction accuracy in practical applications, this extra running time is acceptable.

[0092] S4, based on the strong prediction model output by S3 and the real-time load data output by S1, sets the initial time window through the sliding window prediction method based on Kalman filtering, dynamically adjusts the window size according to the real-time rate of load change, continuously predicts the power demand, and updates the prediction results in real time according to the new observation value; among them, Figure 4 As shown in the figure, the working method of the sliding window prediction method based on Kalman filtering is as follows: by reading the initial values ​​of historical power consumption data and real-time load data, setting the initial size of the sliding window, initializing the state vector x0 and covariance matrix p0 of the Kalman filter, where x0 represents the power demand forecast value at the initial moment, and p0 represents the uncertainty of the initial state; if the current moment is t, the prediction equation of the Kalman filter uses the state vector x0 at the previous moment t-1 And the process noise covariance matrix Q predicts the current state vector x t And the covariance matrix p at the current moment t ; When the real-time rate of load change is higher than or equal to the preset threshold, the sliding window size is reduced through the dynamic window adjustment mechanism; when the real-time rate of load change is lower than the preset threshold, the sliding window size is expanded; at each time step, the Kalman filter-based sliding window prediction method updates the prediction result according to the current observation value and the state of the prediction model through the Kalman filter algorithm; and after each observation update, the new data is integrated into the sliding window through the data fusion in the sliding window, and the oldest data in the window is removed; if the current moment has not reached the end time of the prediction period, the above method is repeated to achieve continuous prediction.

[0093] In a specific embodiment, Kalman filtering is a recursive Bayesian estimation method for optimally estimating the state of a system in the presence of noise, and its core idea is to estimate the state of the system through two steps of prediction and update. The prediction step predicts the state vector and covariance matrix at the current moment based on the state vector and process noise covariance matrix at the previous moment. The update step updates the prediction result based on the current observation value and the state of the prediction model. The key to Kalman filtering is to minimize the root mean square of the prediction error, thereby providing the optimal estimation result.

[0094] The sliding window processes data within a fixed-size window and gradually moves the window to process new data points. In time series forecasting, the sliding window can be dynamically resized to adapt to the rate of change of the data. The main function of the sliding window is to keep the data up-to-date and relevant, ensuring that the model can respond to the latest changes in a timely manner. By dynamically adjusting the window size, the response speed can be improved when the data changes quickly, and the stability of the prediction can be improved when the data changes slowly. When the load change rate is higher than or equal to the preset threshold, the sliding window size is reduced to respond to changes more quickly; when the load change rate is lower than the preset threshold, the sliding window size is increased to improve the smoothness and stability of the prediction. This mechanism enables the model to maintain the best prediction performance under different circumstances. In power consumption forecasting, the state vector includes the current power consumption, the rate of change of power consumption, etc. The covariance matrix represents the uncertainty of the state vector, reflecting the correlation and variance between the various state variables. Through the covariance matrix, the Kalman filter can quantify the uncertainty of the prediction and gradually reduce this uncertainty in the update step.

[0095] Compared with the existing technology, the sliding window prediction method based on Kalman filtering combines the advantages of Kalman filtering algorithm and sliding window technology, and can achieve high-precision prediction of electricity demand. In addition, this method can dynamically adjust the time range of the prediction model according to the real-time rate of load change to capture changes in electricity demand on different time scales. Secondly, this method can update the prediction results in real time and modify the prediction model according to new observations, thereby further improving the accuracy and reliability of the prediction. At the same time, the implementation process of this method is relatively simple, and it is easy to integrate and expand with other intelligent algorithms and technologies to meet more complex load management and deployment needs.

[0096] In actual work, the hardware working environment of the sliding window prediction method based on Kalman filtering mainly includes: high-performance computing servers, smart meter boxes, communication networks and data storage devices;

[0097] Based on these hardware, the sliding window prediction method based on Kalman filtering (Group A) and the traditional time series prediction algorithm (such as ARIMA, Group B) were used for comparative experiments. Among them, the ARIMA algorithm is a classic time series prediction method that predicts future data points by fitting autoregression and moving average models. The experiment was carried out under the same experimental conditions, and each group conducted five experiments, recording the prediction error (such as mean square error MSE) and prediction time of each experiment. The record table is shown in Table 2:

[0098] Table 2 Comparison of sliding window prediction methods based on Kalman filtering

[0099]

[0100] By comparing the experimental records in Table 2, it can be seen that the sliding window prediction method based on Kalman filtering (Group A) is superior to the traditional ARIMA prediction algorithm (Group B) in both prediction error and prediction time. The mean square error MSE of Group A is significantly lower than that of Group B, indicating that its prediction accuracy is higher; at the same time, the prediction time of Group A is also shorter, indicating that its algorithm has higher operating efficiency.

[0101] S5. Based on the prediction results of S4 and the real-time load data output by S1, the load distribution strategy is dynamically adjusted through the adaptive deep deterministic policy gradient algorithm; wherein, the working method of the adaptive deep deterministic policy gradient algorithm is as follows: T1. Construct an adaptive actor network and an adaptive critic network; the adaptive actor network is used to generate a load distribution strategy according to the current load state of the meter box, and the adaptive critic network is used to predict the reward obtained after the load distribution strategy is adopted; T2. Initialize the adaptive actor network, the adaptive critic network and the experience pool, and the experience pool is used to store historical experience; T3. For each time step t, receive the real-time load data l output by S1 t And the prediction result p output by S4 t , combined with the current load state s t , through the deep convolutional neural network to the current load state s t Perform feature extraction to obtain high-dimensional state representation T4, the adaptive actor network is characterized according to the current state and the current policy parameter θ art , generate load distribution strategy α t ; In the process of generating load distribution strategy, the size of exploration noise is adjusted through adaptive exploration mechanism to balance the relationship between exploration and utilization; T5, generated load distribution strategy α t Executed in the meter box through wireless communication protocol, adjust the load distribution; T6, after executing the action, the adaptive deep deterministic policy gradient algorithm receives the reward r t+1and the next state s t+1 , the tuple (s t ,α t ,r t+1 ,s t+1 ) is stored in the experience pool; T7, during the training process, experience is sampled from the experience pool through an adaptive sampling strategy for training, and the adaptive sampling strategy is weighted according to the timestamp and the reward value to improve the sample utilization efficiency; T8, based on the sampled experience, the adaptive critic network is trained according to the current value function parameter θ cri Calculate the target value and update the value function parameter θ based on the gradient descent algorithm cri , the value function parameter θ cri During the update process, the adaptive critic network calculates the gradient of the actor network through the policy gradient function. The adaptive actor network updates the actor network gradient information provided by the adaptive critic network through the adaptive learning rate δ art Update the policy parameters θ art ; The adaptive learning rate is dynamically adjusted according to the historical learning efficiency and the current strategy stability; wherein the formula expression of the strategy gradient function is:

[0102]

[0103] In formula (1), Represents the objective function J of the actor network to the current policy parameter θ art The gradient of E is used to update the parameters of the actor network. dt Indicates the current load status s t The expected value in the experience pool, which stores historical interaction data, including the four-tuple of state, action, reward and next state (s t ,α t ,r t+1 ,s t+1 ); Represents the gradient of the adaptive critic network to action a, which is used to reflect the current state s t Next, the impact of action a on the Q value; Indicates that in state s t The action generated by the actor network U is: artThe gradient is used to reflect the influence of the parameters of the policy network on the generated actions; T9, at the end of each training cycle, the parameters of the target network are updated according to the current network parameters and the soft update coefficient τ through the target network update mechanism; the soft update coefficient τ is dynamically adjusted according to the fluctuations in the training process to balance the difference between the current network and the target network; T10, the iterative cycle T3 to T9 is terminated when the preset training termination condition is reached; the training termination condition includes the number of training rounds and the convergence of the strategy; during the training process, the early stopping mechanism is used to avoid overfitting and insufficient training.

[0104] In a specific embodiment, the adaptive deep deterministic policy gradient combines the advantages of deep neural networks and deterministic policy gradient optimization. The core of the ADDPG algorithm lies in its two key components: the actor network and the critic network. The actor network is responsible for generating actions (i.e., load distribution strategies) based on the current state (i.e., the prediction results of S4 and the real-time load data output by S1), while the critic network evaluates the quality of the action, that is, predicts the future rewards that can be obtained after taking the action (in load management, rewards can be defined as the optimization of indicators such as energy efficiency and load balancing).

[0105] The actor network models the state space through a deep neural network and outputs a deterministic action, which avoids the randomness of strategy exploration in traditional reinforcement learning and improves learning efficiency. The critic network also uses a deep neural network, which receives the current state and the action generated by the actor network as input, and outputs a scalar value, which is the expected return corresponding to the action. By minimizing the prediction error of the critic network (that is, the difference between the expected return and the actual return), the parameters of the actor network can be continuously optimized so that the actions it generates are closer to the optimal strategy. In addition, the ADDPG algorithm also introduces the concept of a target network to stabilize the training process. The target network is a copy of the actor and critic networks, which are updated regularly to slow down fluctuations during training.

[0106] In the intelligent load management and allocation of meter boxes, the application of the ADDPG algorithm can realize the dynamic adjustment of the load distribution strategy. According to the prediction results of S4 and the real-time load data output by S1, the actor network can generate a distribution strategy that adapts to the current load conditions. This strategy not only takes into account the current load demand, but also predicts future load changes, thereby realizing the advance planning and optimization of the load. Through continuous iterative training, the ADDPG algorithm can learn an efficient load distribution strategy that can balance the loads of each meter box, reduce the occurrence of overload and idle phenomena, and improve the energy efficiency and stability of the entire power grid. In addition, due to the adaptability of the ADDPG algorithm, it can automatically adjust the strategy according to changes in the grid structure and load demand without manual intervention, thereby reducing operation and maintenance costs.

[0107] Compared with existing technologies, the ADDPG algorithm models the state space through deep neural networks, can learn complex load distribution strategies, and the training process is efficient and can quickly converge to the optimal solution. Secondly, the ADDPG algorithm has strong adaptive capabilities and can automatically adjust strategies according to changes in grid structure and load demand without manual intervention. By introducing the target network and minimizing the prediction error of the critic network, the ADDPG algorithm can stabilize the training process and avoid strategy fluctuations and overfitting. In addition, the 1ADDPG algorithm can be integrated and expanded with other intelligent algorithms and technologies, such as combined with deep learning, optimization algorithms, etc., to further improve the efficiency and accuracy of load management.

[0108] In practical applications, the hardware working environment of the adaptive deep deterministic policy gradient algorithm mainly includes:

[0109] High-performance computing server, smart meter box system: including data acquisition module and communication module, communication network: including wired and wireless network, data storage device; these hardware environments are respectively applied to the adaptive deep deterministic policy gradient algorithm (Group A) and the traditional rule-based load distribution algorithm (Group B) for comparative experiments; among them, the traditional algorithm Group B adopts a load distribution strategy based on preset rules and thresholds, and manually sets the priority and restriction conditions of load distribution according to the real-time load of the meter box. Each group conducts five experiments, and records the load balance, energy efficiency, and algorithm running time of each experiment. The comparative experiment records are shown in Table 3:

[0110] Table 3 Comparative experimental records of adaptive deep deterministic policy gradient algorithm

[0111]

[0112] From the data table 3, we can see that the adaptive deep deterministic policy gradient algorithm (Group A) is superior to the traditional rule-based load distribution algorithm (Group B) in terms of load balancing, energy efficiency, and algorithm running time. Group A has a higher load balancing degree, better energy efficiency, and a relatively stable algorithm running time.

[0113] S6, real-time monitoring of load changes, if the load change exceeds the preset threshold, the load is allocated through the real-time allocation algorithm; Figure 5 As shown: The working method of the real-time allocation algorithm is: determine whether the load change exceeds the preset threshold through a threshold comparison mechanism. If not, skip the execution step of the real-time allocation algorithm; if exceeded, generate a priority list through a load distribution priority mechanism, and calculate the load adjustment amount through a load balancing algorithm based on the priority list. The load balancing algorithm determines the optimal adjustment amount of each load device through a linear programming model, and then sends the load adjustment amount to the load device through the control protocol Modbus to adjust the power or switch state of the load device.

[0114] In a specific embodiment, the threshold comparison mechanism is a method for determining whether the load change exceeds a preset threshold. By setting a fixed or dynamic threshold, the system can monitor the load change in real time and trigger the corresponding processing mechanism when the load change exceeds the threshold. The selection of the threshold is usually based on historical data and experience to ensure that unnecessary adjustments are not frequently triggered within the normal range, and that timely responses can be made in abnormal situations. The load distribution priority mechanism is used to generate a priority list to determine which load devices are given priority for adjustment when the load change exceeds the threshold. The priority can be determined based on a variety of factors, such as the importance of the device, energy consumption, and frequency of use. The generation of a priority list helps to prioritize the most important devices under limited adjustment resources to ensure the stability and efficiency of the system. The load balancing algorithm determines the optimal adjustment amount of each load device through a linear programming model. Linear programming is a mathematical optimization method used to maximize or minimize a certain objective function under given constraints. In load balancing, the objective function is usually to minimize the degree of load imbalance, and the constraints include the maximum power limit and the minimum power limit of the device. By solving the linear programming model, the optimal adjustment amount of each device can be found to achieve optimal load distribution. Modbus is a communication protocol used for communication between devices in industrial automation systems. It supports multiple transmission modes, including serial communication and Ethernet communication. In load dispatching, Modbus is used to send load adjustment to load devices to adjust the power or switch status of the devices. The openness and versatility of the Modbus protocol make it widely used in industrial applications.

[0115] Compared with the existing technology, the real-time allocation algorithm can respond quickly when the load exceeds the preset threshold by monitoring the load changes in real time, avoiding equipment damage or power waste caused by load overload. Compared with the traditional periodic inspection method, the real-time allocation algorithm has higher response speed and real-time performance. Through the load distribution priority mechanism and load balancing algorithm, the system can generate the optimal load adjustment plan. The priority mechanism ensures that the most important equipment is given priority under limited adjustment resources, while the load balancing algorithm calculates the optimal adjustment amount of each device through a linear programming model to achieve the optimal distribution of the load. This not only improves the efficiency of the system, but also extends the service life of the equipment. Through the Modbus protocol, the system can automatically send the load adjustment amount to the load device to adjust the power or switch state of the device. This reduces the need for manual intervention and improves the automation level and response speed of the system. Compared with the traditional manual adjustment method, automated control greatly reduces the complexity of operation and the risk of human error. Secondly, the real-time allocation algorithm can dynamically adjust the load distribution strategy according to real-time load changes to adapt to different power consumption environments and load modes. This enables the system to respond and adjust quickly when facing sudden load changes, ensuring the stability and security of the system. In addition, through real-time monitoring and immediate adjustment, the system can take timely measures when the load exceeds the preset threshold to avoid equipment damage or power waste caused by overload. This improves the reliability and safety of the system and ensures the continuity and stability of power supply.

[0116] S7. The distributed collaborative dispatching system is used to coordinate the dispatching and management of cross-regional electric meter boxes. The distributed collaborative dispatching system coordinates the dispatching of cross-regional electric meter boxes through the alternating direction multiplier method. The distributed collaborative dispatching system includes a data access module, an optimization dispatching module and a communication coordination module; the data access module obtains the load data, environmental data and equipment status data of the electric meter boxes in each region through the data interface, and fills in missing values, processes outliers and normalizes the data through the data preprocessing method; the optimization and dispatching module coordinates the dispatching of cross-regional electric meter boxes through the alternating direction multiplier method, and outputs the dispatching strategy to the communication coordination module; the communication coordination module includes a distributed communication unit and a state synchronization unit; the distributed communication unit realizes efficient communication between the electric meter boxes in each region through a message queue; the state synchronization unit ensures the synchronization of the status information of the electric meter boxes in each region through heartbeat detection, and performs redundant backup and fault switching through a fault-tolerant mechanism. Furthermore, the working method of the alternating direction multiplier method is: R1. Define the global objective function Among them, r i represents the load distribution of the electric meter box in the i-th area, f(r i) represents the energy consumption cost and load balance degree of the i-th region; N represents the upper limit of the region; the global objective function includes multiple sub-objective functions, each of which corresponds to an electric meter box in a region; R2, through the auxiliary variable z and the Lagrange multiplier γ, the global optimization problem is decomposed into multiple sub-problems, each of which corresponds to an electric meter box in a region, and the decomposition formula is expressed as follows:

[0117]

[0118] In formula (2), z represents an auxiliary variable, which is used to ensure that the load distribution of all regions is consistent at the global level; γ represents the Lagrange multiplier, which is used to deal with global constraints; β represents the penalty parameter, which is used to balance the weight of local optimization and global constraints; R3. In each iteration step, the load distribution r of the meter box in each region is first updated by the local update function i , ensuring that the load distribution of the electric meter boxes in each area is optimized in the current iteration, the formula expression of the local update function is:

[0119]

[0120] In formula (3), l represents the number of iterations; Represents a dual variable, which is used to transfer information during the iteration process; R4, updates the auxiliary variable z through the global update function to ensure that the load distribution of the meter boxes in all regions is consistent at the global level; the formula expression of the global update function is:

[0121]

[0122] In formula (4), z k+1 Represents the updated auxiliary variables; Represents the load distribution of the electric meter box in each area after the update; R5, updates the Lagrange multiplier γ through the dual update function to ensure that the load distribution of the electric meter boxes in all areas is consistent at the global level. The formula expression of the dual update function is:

[0123]

[0124] In formula (5), γ k+1 Represents the updated Lagrange multiplier; R6. During the iterative update process, if the change in the solution of the subproblem is greater than the preset threshold, the penalty parameter β is increased; if in a certain iteration, if the change in the solution of the subproblem is less than the preset threshold, the penalty parameter β is reduced.

[0125] In a specific embodiment, the data interface is a standardized communication protocol for acquiring data from various data sources. Common data interfaces include API, MQTT, Modbus, etc. These interfaces allow the system to obtain real-time data from different types of sensors, devices, and systems to ensure the diversity and integrity of the data. Data preprocessing fills in missing data points through interpolation methods or statistical-based methods to ensure data integrity. By detecting and removing abnormal data points, the influence of outliers on subsequent analysis is prevented. Finally, the data is scaled to the same order of magnitude to facilitate subsequent processing and analysis. Commonly used methods include minimum-maximum normalization and Z-score normalization.

[0126] In the optimization scheduling module, the alternating direction multiplier method ADMM is a distributed algorithm for solving convex optimization problems. It decomposes a complex global optimization problem into multiple easy-to-handle sub-problems and solves them in parallel on multiple computing nodes. By introducing Lagrange multipliers and consistency constraints, ADMM can alternately update variables and multipliers between different nodes until the global optimal solution or approximate optimal solution is reached. This method is particularly suitable for large-scale data sets and distributed computing environments. In the intelligent load management and dispatching of meter boxes, ADMM can decompose the load scheduling problem of cross-regional meter boxes into multiple sub-region problems, each of which is responsible for the load optimization of a part of the meter boxes. By iteratively updating the load distribution strategy and Lagrange multipliers of each sub-region, ADMM can gradually approach the global optimal load distribution solution. This distributed solution method significantly improves the computational efficiency and reduces the dependence on a single computing node. Compared with traditional centralized optimization algorithms, ADMM has better scalability and robustness. It can handle larger-scale power grid data and achieve efficient load scheduling in a distributed computing environment. In addition, the iterative update mechanism of ADMM enables the algorithm to gradually approach the optimal solution, improving the accuracy and stability of load scheduling.

[0127] In the communication coordination module, the distributed communication unit realizes efficient communication between the meter boxes in each area through the message queue. The message queue is an asynchronous communication mechanism that supports high concurrency and low latency. Common message queues include RabbitMQ, Kafka, etc. These tools ensure the reliable delivery of messages and support the scalability and reliability of the system.

[0128] The heartbeat detection mechanism in the state synchronization unit ensures the synchronization of the status information of the meter boxes in each area by sending heartbeat packets regularly. If a node fails to send a heartbeat packet on time, the system will consider that the node may be faulty. The fault tolerance mechanism ensures the high availability and reliability of the system through redundant backup and fault switching. Redundant backup prevents single point failure by saving the same data copies on multiple nodes; after detecting a faulty node, the fault switching automatically switches to the backup node to ensure the continuous operation of the system.

[0129] Compared with the existing technology, the distributed collaborative dispatching system and the alternating direction multiplier method have shown significant advantages and characteristics in the intelligent load management and dispatching method of the meter box. First, the distributed collaborative dispatching system realizes the intelligent and automated management and dispatching of cross-regional meter boxes through the collaborative work of the three modules of data access, optimized dispatching and communication coordination. This greatly improves the energy efficiency and stability of the power grid and reduces operating costs and energy consumption. Secondly, the application of the alternating direction multiplier method enables the system to efficiently solve large-scale and structured optimization problems, and improves the accuracy and feasibility of the dispatching strategy. In addition, the distributed collaborative dispatching system also has good scalability and adaptability, and can flexibly adapt to the needs of power grids of different scales and structures. Finally, the system ensures the information consistency between the meter boxes in each region and the stable operation of the system through an efficient communication mechanism and state synchronization strategy. These advantages and characteristics make the distributed collaborative dispatching system and the alternating direction multiplier method have important application value and development prospects in the intelligent load management and dispatching method of the meter box.

[0130] In the specific implementation, the hardware working environment of the alternating direction multiplier method mainly includes high-performance computing server clusters, data acquisition and transmission equipment (smart meters, sensors, etc.), communication networks (Ethernet, fiber optic networks, etc.) and data storage and processing equipment (database servers, data warehouses, etc.).

[0131] Based on the above hardware, the distributed collaborative scheduling system and the alternating direction multiplier method (Group A) and the traditional algorithm (Group B) were used for comparative experiments. The traditional algorithm uses a centralized optimization algorithm, that is, the load data of all meter boxes are concentrated on a computing node for optimization processing, and then the scheduling strategy is output. During the experiment, Group A and Group B were conducted under the same experimental conditions, and each group conducted five experiments. In order to more professionally evaluate the performance of the two methods, the experiment will record the following parameters:

[0132] Load balancing standard deviation: used to measure the uniformity of load distribution among various meter boxes. The smaller the standard deviation, the higher the load balancing degree.

[0133] Calculate the convergence speed: It indicates the time or number of iterations required for the algorithm to reach a stable state. The faster the convergence speed, the more efficient the algorithm.

[0134] Communication overhead: measures the network bandwidth and latency required for data exchange and information synchronization in a distributed environment. Lower communication overhead indicates higher system resource utilization.

[0135] System robustness: Evaluate the system's ability to recover and stability when facing network failures or node failures. The stronger the robustness, the higher the system reliability. The experimental record table is shown in Table 4:

[0136] Table 4 Alternating direction multiplier method comparison experiment record

[0137]

[0138]

[0139] As shown in Table 4, group A using the distributed collaborative scheduling system and the alternating direction multiplier method is superior to group B using the traditional algorithm in terms of load balancing standard deviation, computational convergence speed, communication overhead, and system robustness. The load balancing standard deviation of group A is significantly lower, indicating that the load is more evenly distributed among the meter boxes; the computational convergence speed is faster, which reduces the time required for the algorithm to reach a stable state; the communication overhead is lower, which improves the utilization of system resources; the system is more robust, which enhances the system's recovery ability and stability when facing network failures or node failures. These results show that the distributed collaborative scheduling system and the alternating direction multiplier method can significantly improve the performance and reliability of the meter box intelligent load management and dispatching system, and provide strong support for the optimized operation of the smart grid.

[0140] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that these specific embodiments are only illustrative, and those skilled in the art may omit, replace, and change the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, merging the above method steps so as to perform substantially the same functions in substantially the same manner to achieve substantially the same results is within the scope of the present invention. Therefore, the scope of the present invention is limited only by the appended claims.

Claims

1. A method for intelligent load management and deployment of an electric meter box, characterized in that: The following steps are involved: S1. Collect and analyze diversified data through real-time data interface and gradient boosting tree, including real-time load data, historical power consumption data, weather forecast, seasonal information, holiday schedule, to form a comprehensive data set; the historical power consumption data includes power consumption in different time periods and under different weather conditions; S2. Based on the comprehensive data set, a power consumption prediction model is constructed through the long short-term memory network LSTM to predict power consumption in real time; S3, based on the historical electricity consumption data in S1, enhances the historical electricity consumption data by synthesizing minority class oversampling, eliminates regional and time period bias, and trains a strong prediction model through an adaptive enhancement algorithm; S4, based on the strong prediction model output by S3 and the real-time load data output by S1, sets the initial time window through the sliding window prediction method based on Kalman filtering, dynamically adjusts the window size according to the real-time rate of load change, continuously predicts the power demand, and updates the prediction results in real time according to the new observation value; S5, based on the prediction results of S4 and the real-time load data output by S1, dynamically adjust the load distribution strategy through the adaptive deep deterministic policy gradient algorithm; S6. Monitor load changes in real time. If the load change exceeds a preset threshold, load allocation is performed through an instant allocation algorithm. S7. Coordinated allocation, scheduling and management of cross-regional electricity meter boxes are performed through a distributed collaborative scheduling system, wherein the distributed collaborative scheduling system performs collaborative scheduling of cross-regional electricity meter boxes through an alternating direction multiplier method.

2. The method for intelligent load management and deployment of an electric meter box according to claim 1, characterized in that: The analysis method of analyzing diversified data by gradient boosting tree in S1 is as follows: first, the real-time load data of the meter box, the historical power consumption data and the diversified information of the external data source are integrated through the real-time data interface, and the integrated data is pre-processed by outlier detection and normalization processing; the external data source includes weather forecast, seasonal information, and holiday schedule; Based on the preprocessed data set, a series of basic decision trees are generated as weak learners, and each weak learner independently performs a preliminary fit on the data; in the iterative process, the residual of the previous round of prediction results is calculated by the gradient descent method, and the training target is updated according to the residual; if the residual is greater than the preset threshold, the weak learner is regenerated; in the new round of iterations, the nodes of the decision tree are split according to the importance of the features through the feature weighting and splitting strategy, and the weights of the nodes are adjusted; after the iteration meets the preset threshold, based on the iteratively generated weak learners, the weak learners generated in each round of iteration are integrated into the strong learner through the ensemble learning method to form a gradient boosting tree model; then the loss value of the current gradient boosting tree model is calculated by the mean square error MSE, and the parameters of the gradient boosting tree model are adjusted according to the loss value; finally, the gradient boosting tree model is regularized by L1 regularization, and the decision tree is pruned.

3. The method for intelligent load management and deployment of an electric meter box according to claim 1, characterized in that: The power consumption prediction model includes an input layer, an embedding layer, an LSTM layer, an attention mechanism layer, a fully connected layer and an output layer; the working method of the power consumption prediction model is: F1. receiving multi-source data in a comprehensive data set through the input layer, and the input layer ensures the format and quality of the data through normalization and standardization methods; F2, converting discrete category data into continuous vector representations through the embedding layer; the embedding layer maps each category to a fixed-length vector through a lookup table to capture the characteristics of the category data, the category data including holiday schedules and seasonal information; F3, capturing long-term dependencies in time series data through the LSTM layer, which controls the inflow and outflow of information through input gates, forget gates, and output gates, dynamically adjusts the degree of memory of historical information, and captures long-term dependencies in time series data through cell states; F4, highlighting important time steps and features through the attention mechanism layer; The attention mechanism layer generates an attention map by calculating the weight of each time step and feature; F5. The outputs of the LSTM layer and the attention mechanism layer are integrated through the fully connected layer to generate a high-dimensional feature vector; F6. Generate a power consumption prediction result through the output layer; the output layer performs a nonlinear transformation on the high-dimensional feature vector through the activation function ReLU, and converts the prediction result into an actual power consumption value through a normalization operation.

4. The method for intelligent load management and deployment of an electric meter box according to claim 1, characterized in that: The method for enhancing the historical electricity consumption data by synthetic minority oversampling in S3 is as follows: first, the historical electricity consumption data is preliminarily divided by a density-based clustering algorithm to identify electricity consumption patterns in different regions and time periods, including minority samples and all samples; the minority samples represent data points in time periods and regions with low or high electricity consumption, and based on the identified minority samples, the Euclidean distance between the minority samples and all other samples is calculated by a nearest neighbor-based interpolation method to find the K nearest neighbor samples of the minority samples; according to the distribution of the K nearest neighbor samples, one of them is selected for interpolation to generate a new minority sample.

5. The method for intelligent load management and deployment of an electric meter box according to claim 1, characterized in that: The working method of the adaptive enhancement algorithm is as follows: based on the enhanced data set, the adaptive enhancement algorithm first assigns weights to each sample in the data set through an initial weight assignment method, then trains the first weak prediction model based on the current weight distribution, and calculates the prediction error of the weak prediction model on the training set; according to the prediction error of the weak prediction model, calculates the error rate of each sample; according to the error rate, dynamically adjusts the weight of each sample through a weight adjustment strategy, focuses on fuzzy samples, and performs multiple iterative training; During the iteration process, the adaptive enhancement algorithm dynamically adjusts the weight of the next round of weak prediction models according to the prediction error of each weak prediction model on the training set, and weightedly combines the models after each round of adjustment to generate a strong prediction model; during each iterative training process, the adaptive enhancement algorithm also dynamically adjusts the weight of the input features according to the feature importance score of the weak prediction model through the feature importance evaluation mechanism.

6. The method for intelligent load management and deployment of an electric meter box according to claim 1, characterized in that: The working method of the sliding window prediction method based on Kalman filtering is as follows: by reading the initial values ​​of historical power consumption data and real-time load data, setting the initial size of the sliding window, initializing the state vector x0 and covariance matrix p0 of the Kalman filter, wherein x0 represents the power demand forecast value at the initial moment, and p0 represents the uncertainty of the initial state; if the current moment is t, then the state vector x0 at the previous moment is used to predict the power demand at the previous moment through the prediction equation of the Kalman filter. t-1 And the process noise covariance matrix Q predicts the current state vector x t And the covariance matrix p at the current moment t ; When the real-time rate of load change is higher than or equal to the preset threshold, the sliding window size is reduced through the dynamic window adjustment mechanism; when the real-time rate of load change is lower than the preset threshold, the sliding window size is expanded; at each time step, the Kalman filter-based sliding window prediction method updates the prediction result according to the current observation value and the state of the prediction model through the Kalman filter algorithm; and after each observation update, the new data is integrated into the sliding window through the data fusion in the sliding window, and the oldest data in the window is removed; if the current moment has not reached the end time of the prediction period, the above method is repeated to achieve continuous prediction.

7. The method for intelligent load management and deployment of an electric meter box according to claim 1, characterized in that: The working method of the adaptive deep deterministic policy gradient algorithm is: T1. Construct an adaptive actor network and an adaptive critic network; the adaptive actor network is used to generate a load distribution strategy according to the current load state of the meter box, and the adaptive critic network is used to predict the reward obtained after adopting the load distribution strategy; T2, initializing the adaptive actor network, the adaptive critic network and the experience pool, wherein the experience pool is used to store historical experience; T3, for each time step t, receiving the real-time load data l output by S1 t And the prediction result p output by S4 t , combined with the current load state s t , through the deep convolutional neural network to the current load state s t Perform feature extraction to obtain high-dimensional state representation T4, the adaptive actor network is characterized according to the current state and the current policy parameter θ art , generate load distribution strategy α t ;In the process of generating load distribution strategy, the size of exploration noise is adjusted through adaptive exploration mechanism to balance the relationship between exploration and utilization; T5, generated load distribution strategy α t Adjust load distribution through wireless communication protocol executed on the meter box; T6. After executing the action, the adaptive deep deterministic policy gradient algorithm receives the reward r t+1 and the next state s t+1 , the tuple (s t ,α t ,r t+1 ,s t+1 ) is stored in the experience pool; T7. During the training process, experience is sampled from the experience pool through an adaptive sampling strategy for training. The adaptive sampling strategy is weighted according to the timestamp and the reward value to improve the sample utilization efficiency; T8. Based on the sampled experience, the adaptive critic network adjusts the value function parameters θ according to the current value function parameters θ. cri Calculate the target value and update the value function parameter θ based on the gradient descent algorithm cri , the value function parameter θ cri During the update process, the adaptive critic network calculates the gradient of the actor network through the policy gradient function. The adaptive actor network updates the actor network gradient information provided by the adaptive critic network through the adaptive learning rate δ art Update the policy parameters θ art ; The adaptive learning rate is dynamically adjusted according to the historical learning efficiency and the current strategy stability; wherein the formula expression of the strategy gradient function is: In formula (1), Represents the objective function J of the actor network to the current policy parameter θ art The gradient of E is used to update the parameters of the actor network. dt Indicates the current load status s t The expected value in the experience pool, which stores historical interaction data, including the four-tuple of state, action, reward and next state (s t ,α t ,r t+1 ,s t+1 ); Represents the gradient of the adaptive critic network to action a, which is used to reflect the current state s t Next, the impact of action a on the Q value; Indicates that in state s t The action generated by the actor network U is: art The gradient of is used to reflect the influence of the parameters of the policy network on the generated actions; T9. At the end of each training cycle, the parameters of the target network are updated according to the current network parameters and the soft update coefficient τ through the target network update mechanism; the soft update coefficient τ is dynamically adjusted according to the fluctuations during the training process to balance the difference between the current network and the target network; T10, iterative loop T3 to T9, terminated when the preset training termination conditions are reached; the training termination conditions include the number of training rounds and strategy convergence; during the training process, the early stopping mechanism is used to avoid overfitting and insufficient training.

8. The method for intelligent load management and deployment of an electric meter box according to claim 1, characterized in that: The working method of the instant allocation algorithm is: judging whether the load change exceeds a preset threshold through a threshold comparison mechanism, and if not, skipping the execution step of the instant allocation algorithm; If it exceeds the limit, a priority list is generated through the load distribution priority mechanism, and the load adjustment amount is calculated based on the priority list through the load balancing algorithm. The load balancing algorithm determines the optimal adjustment amount for each load device through a linear programming model, and then sends the load adjustment amount to the load device through the control protocol Modbus to adjust the power or switching state of the load device.

9. The method for intelligent load management and deployment of an electric meter box according to claim 1, characterized in that: The distributed collaborative scheduling system includes a data access module, an optimization scheduling module and a communication coordination module; the data access module obtains the load data, environmental data and equipment status data of the meter boxes in each area through the data interface, and fills the missing values, processes the abnormal values ​​and normalizes the data through the data preprocessing method; the optimization and scheduling module performs collaborative scheduling of cross-regional meter boxes through the alternating direction multiplier method, and outputs the scheduling strategy to the communication coordination module; the communication coordination module includes a distributed communication unit and a state synchronization unit; the distributed communication unit realizes efficient communication between the meter boxes in each area through a message queue; the state synchronization unit ensures the synchronization of the status information of the meter boxes in each area through heartbeat detection, and performs redundant backup and fault switching through a fault-tolerant mechanism.

10. The method for intelligent load management and deployment of an electric meter box according to claim 9, characterized in that: The alternating direction multiplier method works as follows: R1. Define the global objective function Among them, r i represents the load distribution of the electric meter box in the i-th area, f(r i ) represents the energy consumption cost and load balance degree of the i-th area; N represents the upper limit of the area; the global objective function includes multiple sub-objective functions, each sub-objective function corresponds to an electric meter box in an area; R2. Decompose the global optimization problem into multiple sub-problems through auxiliary variables z and Lagrange multiplier γ. Each sub-problem corresponds to the meter box in a region. The decomposition formula is: In formula (2), z represents the auxiliary variable, which is used to ensure that the load distribution of all regions is consistent at the global level; γ represents the Lagrange multiplier, which is used to deal with global constraints; β represents the penalty parameter, which is used to balance the weight of local optimization and global constraints; R3. In each iteration step, the load distribution of the meter box in each area is first updated by the local update function r i , ensuring that the load distribution of the electric meter boxes in each area is optimized in the current iteration, the formula expression of the local update function is: In formula (3), k represents the number of iterations; Represents the dual variable, which is used to transmit information during the iteration process; R4. Update the auxiliary variable z through the global update function to ensure that the load distribution of the meter boxes in all regions is consistent at the global level; the formula expression of the global update function is: In formula (4), z k+1 Represents the updated auxiliary variables; Indicates the updated load distribution of the electric meter boxes in each area; R5. Update the Lagrange multiplier γ through the dual update function to ensure that the load distribution of the meter boxes in all regions is consistent at the global level. The formula expression of the dual update function is: In formula (5), γ k+1 represents the updated Lagrange multiplier; R6. During the iterative update process, if the change in the solution of the sub-problem is greater than the preset threshold, the penalty parameter β is increased; if in a certain iteration, if the change in the solution of the sub-problem is less than the preset threshold, the penalty parameter β is reduced.

Images