Data monitoring method and system for electricity marketing

Abstract

The invention provides a data monitoring method and system for power marketing. The method comprises the following steps: collecting original data from a plurality of power marketing data sources; constructing a space-time diagram, and predicting electric power by using a neural network model according to the global space-time feature vector; performing dynamic strategy optimization by using a deep reinforcement learning model according to a prediction result of the neural network model; collecting feedback data after strategy optimization is executed in real time, designing a strategy evaluation function to evaluate the effect of the marketing strategy, and correcting the deep reinforcement learning model according to an evaluation result; and according to real-time feedback data, the optimized strategy is applied to different users according to personalized requirements of different users, and it is guaranteed that each user obtains recommendation most meeting the requirements of the user. According to the invention, the demand fluctuation of the electricity market can be captured in real time, and the electricity marketing strategy is adjusted according to the demand changes of different user groups, so that the accuracy, flexibility and effect of electricity marketing are greatly improved.

Description

Technical Field

[0001] The present invention belongs to the field of data monitoring, and in particular relates to a data monitoring method and system for power marketing. Background Art

[0002] With the gradual reform of the global power market and the rapid development of smart grid technology, power marketing is no longer based solely on traditional fixed prices and packages, but is moving towards a more flexible, personalized and precise direction. In this process, power companies have begun to rely on a large amount of real-time data, trying to optimize marketing strategies through accurate user portraits and real-time monitoring to improve user satisfaction and corporate economic benefits. Specifically, power companies need to adjust power marketing strategies in real time based on various factors such as the needs of different users, electricity consumption behavior, weather conditions, holidays, etc., and make corresponding adjustments and optimizations based on user feedback.

[0003] However, most current power marketing systems still have some technical limitations, which make it difficult for them to achieve true personalization, real-time and dynamic. First, existing power marketing systems often rely on traditional statistical analysis methods or static data models, and are unable to capture the rapid changes in user needs in real time. For example, many power marketing systems make predictions and recommendations based on past electricity consumption data, but these methods are insufficient in dealing with emergencies (such as weather changes, holiday effects, etc.). More importantly, existing systems find it difficult to effectively integrate multi-source heterogeneous data, and can often only process structured historical data, ignoring the potential value from unstructured data sources (such as social media, user feedback, etc.).

[0004] In addition, traditional power marketing decisions usually lack a real-time feedback mechanism, resulting in a lack of flexibility in adjusting marketing strategies after they are implemented. Once a marketing strategy is not effective, it is difficult for the existing system to make effective adjustments or optimizations in a short period of time. The response mechanism of many systems is relatively lagging, and users’ dissatisfaction with a certain power package or service often cannot be quickly fed back to the system, resulting in unsatisfactory marketing results and even affecting users’ overall satisfaction with the power company.

[0005] In view of the shortcomings of these existing technologies, the power marketing system urgently needs a solution that can process multi-source heterogeneous data, adjust decisions in real time, and quickly optimize marketing strategies based on user feedback. Especially in the power market, how to deal with the drastic fluctuations in market demand, how to process massive amounts of real-time data, and how to effectively combine user personalized needs with market dynamics have become key challenges for power companies to improve their competitiveness and user experience. Summary of the invention

[0006] The purpose of the present invention is to propose a data monitoring method and system for power marketing, which solves the deficiencies of the existing power marketing system in terms of real-time performance, data fusion and dynamic adjustment of decision-making.

[0007] In order to achieve the above object, a first aspect of the present invention provides a data monitoring method for power marketing, the method comprising:

[0008] S1. Collect raw data from multiple power marketing data sources, standardize and preprocess the raw data, and fuse the data using weighted aggregation after preprocessing to obtain weighted fused features. t Integrate into a unified feature representation X to form a data set X = {X1,X2,…,X t ,X T};

[0009] S2. Constructing a spatiotemporal graph based on multimodal features, using a graph convolutional neural network to update node features of the spatiotemporal graph, and introducing a spatiotemporal dependency regularization term during the updating process to improve the performance of the spatiotemporal graph convolutional network in power marketing prediction, and then aggregating all updated nodes to obtain a global spatiotemporal feature vector, and using a neural network model to predict power based on the global spatiotemporal feature vector;

[0010] S3. Use the deep reinforcement learning model to perform dynamic strategy optimization based on the prediction results of the neural network model, including:

[0011] The global spatiotemporal feature vector is used as the state space input of reinforcement learning. The state space contains the global spatiotemporal features at each time t. The action space is defined as the operation performed by the agent at each time step. The reward function r t Including demand error term and energy cost term, it is expressed as follows:

[0012]

[0013] in, is the predicted electricity demand, y t It is a practical need. Is to perform an action The energy cost at this time, λ1 and λ2 are the weight coefficients of control error and cost respectively, i is the node i, and N is the total number of nodes;

[0014] Build a deep reinforcement learning model, including:

[0015] (a) Using DQN to optimize the power dispatch strategy, where the update rule of the DQN is expressed as follows:

[0016]

[0017] Among them, α is the learning rate, γ is the discount factor, and r t It's a reward. The next state The maximum Q value of all actions; is the state space, i.e. the global space-time eigenvector; a t It is the optimization result of scheduling strategy;

[0018] (b) A deep reinforcement learning model consisting of a two-layer fully connected neural network is used to approximate the Q-value function. The network structure is:

[0019]

[0020] Among them, NN is a neural network, and W represents the weight in the network. The input of the neural network is the spatiotemporal features The output is the Q value of different actions;

[0021] The time-space dependency regularization term is introduced to constrain power dispatching to prevent the intelligent agent from making incoherent dispatching decisions. At the same time, the standard network and experience replay technology are used to avoid the oscillation of Q value in the deep reinforcement learning process to ensure the stability of the reinforcement learning model.

[0022] Design a feedback mechanism to dynamically adjust the power dispatch strategy by monitoring the difference between power demand and predicted demand in real time. After each decision, the agent fine-tunes the strategy based on the error between actual demand and predicted demand.

[0023] S4. Collect feedback data after execution strategy optimization in real time, design a strategy evaluation function to evaluate the effect of the marketing strategy, and modify the deep reinforcement learning model according to the evaluation results; the feedback data includes actual power demand, system response status and predicted power demand;

[0024] S5. Based on real-time feedback data, the optimized strategy is applied to different users according to their personalized needs to ensure that each user receives the recommendation that best meets their needs.

[0025] Furthermore, the standardization process is expressed as:

[0026] Assume x ij represents the original data point, then the standardized data Calculated by the following formula:

[0027]

[0028] in, is the data point x ij The mean of is the data point xij The standard deviation of

[0029] The preprocessing is expressed as:

[0030] Set a unified timestamp set t1, t2, ..., t T , and use the interpolation method to convert the data x in each data source ij and timestamp t j Corresponding to these unified timestamps, we get the interpolated data

[0031]

[0032] The weighted aggregation method is used to fuse the data to obtain the weighted fused features, which are expressed as:

[0033] Assume that each data source The standardized data is Weighted fusion feature X t At timestamp t, it is calculated by the following formula:

[0034]

[0035] Among them, w i Represents the data source The weight of Is the data source Normalized data at timestamp t.

[0036] Furthermore, the space-time graph is constructed as follows:

[0037] Define an adjacency matrix A t To represent the spatiotemporal association between different power users at timestamp t, there are N power users and the adjacency matrix A t Element A ij represents the spatiotemporal similarity between user i and user j at timestamp t, which is calculated as follows:

[0038]

[0039] Among them, X i and X j are the feature vectors of user i and user j at timestamp t, σ s is the spatial similarity control parameter, represents the neighbor set of user i;

[0040] The node features of the spatiotemporal graph are updated using a graph convolutional neural network, and the graph convolution operation is as follows:

[0041]

[0042] Among them, W t is the weight matrix at timestamp t, σ is the activation function, X′ i (t) is the updated feature;

[0043] The spatiotemporal dependency regularization term R t , expressed as:

[0044]

[0045] Among them, λ is a hyperparameter that controls the regularization strength, A ij is the weight of the spatiotemporal association;

[0046] The aggregation of all updated nodes is expressed as:

[0047]

[0048] in, is the aggregated spatiotemporal feature vector, and N is the number of nodes in the graph;

[0049] The power is predicted using a neural network model according to the global spatiotemporal feature vector, which is represented by:

[0050]

[0051] in, is the predicted electricity demand, W p and b p are the weight matrix and bias term of the neural network, and σ is the activation function.

[0052] Further, when obtaining the predicted power demand Then compare it with the actual power demand t Compare to calculate the error, and optimize the neural network model according to the error, which is expressed as:

[0053]

[0054] At the same time, according to the error Perform online learning and update the model parameters of the neural network model through the adaptive gradient descent algorithm, which is expressed as:

[0055]

[0056] Where η is the learning rate, is the gradient of the loss function with respect to the weights, G t is the sum of squares of historical gradients, and ∈ is a small constant to prevent division by zero errors.

[0057] Furthermore, the spatiotemporal regularization term is constructed as follows:

[0058] Assume that the adjacency matrix of user i and user j in the space-time graph is A ij , the spatiotemporal regularization term The temporal and spatial consistency of decisions is improved by encouraging the scheduling strategy differences between adjacent users to be minimized, which can be expressed as:

[0059]

[0060] Among them, λ3 is the regularization hyperparameter, A ij is the similarity between user i and user j, and Respectively represent the scheduling actions of user i and user j at time step t;

[0061] The Q value update rule of the target network is as follows:

[0062]

[0063] Among them, Q′ represents the target network, Q represents the current network;

[0064] The experience replay is that after each interaction with the environment, the agent stores the current state, action, reward and next state into the experience pool;

[0065] The feedback mechanism is designed as follows:

[0066] Assume that at time step t, the scheduling decision chosen by the agent is a t , the actual power demand is y t , the forecast demand is The feedback-based strategy update is expressed as:

[0067]

[0068] Where ΔW is the weight update, η is the learning rate, is the gradient of the Q value, and t Forecast and actual demand respectively.

[0069] Furthermore, the S4 specifically includes:

[0070] Calculate the dispatch error based on the actual and predicted power demand, calculate the system stability based on the system response state, and design the reward function R based on the dispatch error and system stability t Feedback to the agent for dynamic strategy adjustment and retraining; wherein the scheduling error E t The calculation is as follows:

[0071]

[0072] The system stability S t The calculation is as follows:

[0073]

[0074] in, is the i-th dimension of the system state, and M is the system state dimension;

[0075] The reward function R t , expressed as:

[0076] R t =-λ1E t -λ2S t

[0077] Among them, λ1,λ2 are the weighted coefficients of scheduling error and system stability in the reward function;

[0078] Design the current policy evaluation function P t , quantify the overall effect of the current scheduling strategy, expressed as:

[0079]

[0080] Among them, α1, α2, and α3 are hyperparameters used to balance scheduling error, system stability, and total scheduling. is the power allocation of the agent to the i-th user;

[0081] Training auxiliary models Used to predict the long-term response of the system under different scheduling strategies:

[0082]

[0083] in, It is an auxiliary model obtained through training, which is used to predict the impact of scheduling decisions on the long-term stability of the system.

[0084] Furthermore, the design reward function R t Feedback is given to the agent for dynamic strategy adjustment and retraining, specifically:

[0085] Update the Q value function using the TemporalDifference method:

[0086]

[0087] Among them, α is the learning rate, R t It's a reward.

[0088] The nonlinear regularization term is introduced to enhance the stability of the system, and the modified Q value update formula is as follows:

[0089]

[0090] Among them, β is the regularization coefficient.

[0091] Furthermore, the S5 specifically includes:

[0092] Collect user needs, combine time series analysis, and build personalized demand vectors And through the clustering algorithm, the user group is divided into different demand groups; the user demand includes historical power load, response mode and preference;

[0093] A personalized recommendation strategy model is built using a deep neural network to generate personalized scheduling decisions based on user demand vectors and system status;

[0094] Design a comprehensive utility function to evaluate personalized scheduling decisions;

[0095] Whenever the system state or user feedback changes, the personalized scheduling decision is updated by calculating real-time rewards to evaluate the effectiveness of the current strategy;

[0096] Combine personalized dispatch decisions with global dispatch strategies to coordinate and optimize power marketing data;

[0097] At the end of each cycle, the long-term benefits of personalized scheduling decisions are evaluated and the strategy is updated based on continuous feedback from users. At the same time, whenever the global scheduling strategy encounters an emergency, the personalized scheduling decision needs to be adjusted in a timely manner.

[0098] Furthermore, the comprehensive utility function is expressed as:

[0099]

[0100] in, is the load satisfaction of user i at time t; is the response efficiency, which takes into account how quickly the user responds to the dispatch within the specified time; is the electricity cost saved by the user, which measures the impact of the dispatch strategy on the user's economic benefits; γ1, γ2, and γ3 are weight coefficients, reflecting the relative importance of different factors to the user.

[0101] In another aspect of the present invention, a data monitoring system for power marketing is provided, the system comprising:

[0102] The power marketing data acquisition unit is used to collect raw data from multiple power marketing data sources, standardize and preprocess the raw data, and fuse the data using weighted aggregation after preprocessing to obtain weighted fused features. The multimodal feature X of each timestamp t is t Integrate into a unified feature representation X to form a data set X = {X1,X2,…,X t ,X T};

[0103] A space-time graph construction unit is used to construct a space-time graph, update node features of the space-time graph using a graph convolutional neural network, introduce a space-time dependency regularization term during the update process to improve the performance of the space-time graph convolutional network in power marketing prediction, and then aggregate all updated nodes to obtain a global space-time feature vector, and use a neural network model to predict power based on the global space-time feature vector;

[0104] The deep reinforcement learning unit is used to perform dynamic strategy optimization using the deep reinforcement learning model based on the prediction results of the neural network model, including:

[0105] The global spatiotemporal feature vector is used as the state space input of reinforcement learning. The state space contains the global spatiotemporal features at each time t. The action space is defined as the operation performed by the agent at each time step. The reward function r t Including demand error term and energy cost term, it is expressed as follows:

[0106]

[0107] in, is the predicted electricity demand, y t It is a practical need. Is to perform an action The energy cost at this time, λ1 and λ2 are the weight coefficients of control error and cost respectively, i is the node i, and N is the total number of nodes;

[0108] Build a deep reinforcement learning model, including:

[0109] (a) Using DQN to optimize the power dispatch strategy, where the update rule of the DQN is expressed as follows:

[0110]

[0111] Among them, α is the learning rate, γ is the discount factor, and r t It's a reward. The next state The maximum Q value of all actions; is the state space, i.e. the global space-time eigenvector; at It is the optimization result of scheduling strategy;

[0112] (b) A deep reinforcement learning model consisting of a two-layer fully connected neural network is used to approximate the Q-value function. The network structure is:

[0113]

[0114] Among them, NN is a neural network, and W represents the weight in the network. The input of the neural network is the spatiotemporal features The output is the Q value of different actions;

[0115] The time-space dependency regularization term is introduced to constrain power dispatching to prevent the intelligent agent from making incoherent dispatching decisions. At the same time, a two-layer fully connected neural network and experience replay technology are used to avoid Q-value oscillation during deep reinforcement learning and ensure the stability of the reinforcement learning model.

[0116] Design a feedback mechanism to dynamically adjust the power dispatch strategy by monitoring the difference between power demand and predicted demand in real time. After each decision, the agent fine-tunes the strategy based on the error between actual demand and predicted demand.

[0117] A real-time optimization unit is used to collect feedback data after executing strategy optimization in real time, design a strategy evaluation function to evaluate the effect of the marketing strategy, and modify the deep reinforcement learning model according to the evaluation results; the feedback data includes actual power demand, system response status and predicted power demand;

[0118] The personalized recommendation unit is used to apply the optimized strategy to different users according to their personalized needs based on real-time feedback data, ensuring that each user receives the recommendation that best meets their needs.

[0119] The beneficial technical effects of the present invention are at least as follows:

[0120] (1) The present invention fuses the heterogeneous data required for power marketing by introducing a spatiotemporal graph convolutional network (ST-GCN). Unlike existing systems that rely only on structured data (such as historical electricity consumption data), the present invention not only considers structured data, but also includes unstructured data such as weather, holidays, regional economic activities, and social media feedback in the analysis scope. Through spatiotemporal graph convolutional network processing, these multi-dimensional, heterogeneous data can be effectively integrated, thereby providing strong data support for the precise formulation of power marketing strategies. Through the modeling of spatiotemporal data, the system can identify the impact of various environmental factors on user needs and adjust marketing strategies in real time.

[0121] (2) In order to make up for the lag problem in strategy adjustment of the existing power marketing system, the present invention introduces a deep reinforcement learning (DRL) algorithm, which enables the power marketing system to dynamically optimize marketing strategies based on real-time data. By combining real-time feedback data, the system can automatically adjust power packages, preferential plans and recommended content to improve user response rate and conversion rate. For example, when a marketing strategy fails to obtain the expected user feedback, the system can adjust the strategy in real time through a reinforcement learning algorithm and select a promotion method that better meets the current user needs. The introduction of the deep reinforcement learning algorithm enables the power marketing system of the present invention to have a strong adaptive ability and to continuously optimize decisions in a changing market environment.

[0122] (3) The present invention also designs a real-time feedback and adaptive adjustment mechanism. By continuously monitoring user behavior feedback (such as click-through rate, purchase conversion rate, social media evaluation, etc.), the system can evaluate the effectiveness of each marketing activity in real time and dynamically adjust the strategy during the activity. The combination of this feedback mechanism and deep reinforcement learning enables the marketing system to not only make accurate preliminary recommendations, but also make flexible adjustments according to user behavior during the execution process, ensuring that the marketing strategy is always synchronized with user needs and market changes, thereby improving the overall effect of the marketing activities.

[0123] (4) The present invention overcomes the deficiencies of traditional power marketing systems in data processing, real-time decision-making, strategy optimization and flexible adjustment through the organic combination of spatiotemporal data fusion, deep reinforcement learning and real-time feedback mechanism, and provides an innovative and efficient power marketing solution. This system can capture the demand fluctuations in the power market in real time and adjust the power marketing strategy according to the demand changes of different user groups, thereby greatly improving the accuracy, flexibility and effectiveness of power marketing. BRIEF DESCRIPTION OF THE DRAWINGS

[0124] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0125] Figure 1 The present invention is a flow chart of a data monitoring method for power marketing according to an embodiment of the present invention.

[0126] Figure 2 This is a framework diagram of a data monitoring system for power marketing according to an embodiment of the present invention. DETAILED DESCRIPTION

[0127] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0128] like Figure 1 As shown, an embodiment of the present invention provides a data monitoring method for power marketing, the method comprising the following steps S1-S5:

[0129] S1. Collect raw data from multiple power marketing data sources, standardize and preprocess the raw data, and fuse the data using weighted aggregation after preprocessing to obtain weighted fused features. t Integrate into a unified feature representation X to form a data set X = {X1,X2,…,X t ,X T}.

[0130] Specifically, firstly, raw data is collected from multiple power marketing data sources (such as user behavior data, power consumption data, meteorological data, etc.). Assume that the present invention has multiple data sources Each data source provides different types of information, such as users’ electricity usage habits, climate change, social activities, etc. Contains a series of data points x ij , where x ij Is the data source The jth data point in .

[0131] For these data, standardization is required to eliminate the dimensional differences between different data sources so that all feature data are on the same scale. ij represents the original data point, then the standardized data Calculated by the following formula:

[0132]

[0133] in, is the data point x ij The mean of is the data point x ij The standard deviation of the standardized data Will serve as input for subsequent steps.

[0134] Furthermore, the data in the data sources usually have different timestamps, resulting in inconsistent timeliness and synchronization of the data. Therefore, it is necessary to time align and interpolate the data in different data sources. Set a unified timestamp set t1, t2, …, tT , and convert the data x in each data source ij and timestamp t j Corresponding to these unified timestamps.

[0135] By using an interpolation method (such as linear interpolation or spline interpolation), the present invention converts the original data points x of each data source into ij Map to the unified timestamp t to get the interpolated data

[0136]

[0137] This step ensures that the data from all data sources have corresponding timestamps within the same time range and lays the foundation for subsequent multimodal data fusion.

[0138] Furthermore, after data collection and alignment are completed, the data from different data sources need to be fused. According to the importance of each data source, the present invention adopts a weighted aggregation method to fuse the data. Assume that each data source The standardized data is Weighted fusion feature X t At timestamp t, it is calculated by the following formula:

[0139]

[0140] Among them, w i Represents the data source The weight of Is the data source Normalized data at timestamp t. Weight w i Set each data source based on its importance in the power marketing strategy.

[0141] After further weighted aggregation, the obtained X t is a vector containing multimodal features. At this stage, the ultimate goal is to integrate the multimodal features of each timestamp t into a unified feature representation X, forming a dataset X = {X1, X2, …, X T}, where each X t is a multimodal vector containing spatiotemporal features. The dataset X will be used for subsequent spatiotemporal modeling and demand forecasting tasks.

[0142] Finally, the dataset X provided by this step will be used as the input of the spatiotemporal graph convolutional network (ST-GCN) model. Each feature vector X in the dataset tIt contains information from different data sources, such as user behavior, meteorological conditions, social activities, etc. By modeling these multimodal features, more accurate demand forecasting and dynamic decision support can be provided for the power marketing system.

[0143] On this basis, the power marketing system can perform tasks such as demand forecasting and load forecasting through spatiotemporal modeling, thereby improving the accuracy and effectiveness of marketing strategies.

[0144] S2. Construct a spatiotemporal graph based on multimodal features, use a graph convolutional neural network to update the node features of the spatiotemporal graph, and introduce a spatiotemporal dependency regularization term during the update process to improve the performance of the spatiotemporal graph convolutional network in power marketing prediction. Then, aggregate all updated nodes to obtain a global spatiotemporal feature vector, and use a neural network model to predict power based on the global spatiotemporal feature vector.

[0145] Specifically, in this step, a spatiotemporal graph is first constructed based on the spatial distribution and time sequence of power users. t Corresponding to a node in a graph, where the features of each node are represented by X t The edges in the graph represent the spatial dependencies between users (such as adjacent users or users in the same area), and the weights of the edges are defined by the spatiotemporal correlations of these users.

[0146] Furthermore, define an adjacency matrix A t To represent the spatiotemporal association between different power users at timestamp t. Assuming that there are N power users in the present invention, the adjacency matrix A t Element A ij represents the spatiotemporal similarity between user i and user j at timestamp t. At this time, the present invention can use the following formula to calculate the similarity:

[0147]

[0148] Among them, X i and X j are the feature vectors of user i and user j at timestamp t, σ s is the spatial similarity control parameter, represents the neighbor set of user i. This formula calculates the spatiotemporal similarity between users through the Gaussian kernel function and uses normalization to ensure the sparsity of the similarity matrix.

[0149] In this way, the present invention establishes a graph structure with a spatiotemporal relationship, and the node features in the graph are represented by X t Provided, the edge weights are dynamically adjusted according to the spatiotemporal similarity.

[0150] Furthermore, the input of this step is the adjacency matrix A obtained from step 1 t And the spatiotemporal characteristics X t In the spatiotemporal graph convolution operation, the present invention uses a graph convolutional neural network (GCN) to update node features and dynamically adjusts them considering the temporal nature of the spatiotemporal graph.

[0151] Graph convolution operation: The core of spatiotemporal graph convolution is to update the features of nodes based on the structure of the graph. For node i at timestamp t, its updated feature X′ i (t) is calculated by the following formula:

[0152]

[0153] Among them, W t is the weight matrix at timestamp t, and σ is the activation function (such as ReLU). This formula shows that the updated feature of node i is obtained by graph convolution operation using the adjacency matrix A t The features of adjacent nodes are weighted and aggregated to capture the spatial dependency of spatiotemporal data.

[0154] Furthermore, in order to better model the spatiotemporal dependency, the present invention introduces an innovative spatiotemporal dependency

[0155] Dependent regularization term. Assume X i (t) is the feature vector of node i at timestamp t, X′ i (t) is the updated feature, and the spatiotemporal dependency regularization term R t It can be expressed as:

[0156]

[0157] Among them, λ is a hyperparameter that controls the regularization strength, A ij is the weight of spatiotemporal association. The role of this regularization term is to promote the updated node feature X′ i (t) and adjacent nodes X′ j The spatial and temporal differences between (t) are minimized, thereby enhancing the model’s ability to learn spatial and temporal dependencies.

[0158] By introducing spatiotemporal dependency regularization, the present invention effectively improves the performance of spatiotemporal graph convolutional networks in power marketing forecasting, especially when dealing with dynamically changing power demand.

[0159] Furthermore, after the graph convolution operation, we get the updated feature X′ of each node at timestamp t. i (t), these features integrate the spatiotemporal information. Next, the present invention aggregates the features of all nodes to obtain a global spatiotemporal feature vector, and uses the feature to predict power demand.

[0160] Furthermore, in order to obtain global spatiotemporal features, the present invention uses a global average pooling method to aggregate the features of all nodes. Assume X′ i (t) is the updated feature of node i at timestamp t, and the aggregation formula of node features is:

[0161]

[0162] in, is the aggregated spatiotemporal feature vector, and N is the number of nodes in the graph. Through this aggregation operation, the present invention obtains a global feature vector containing spatiotemporal information, which can reflect the demand changes of all users in power marketing.

[0163] Furthermore, the aggregated global spatiotemporal features Input into a neural network model to predict electricity demand. is the spatiotemporal feature vector, the output of the demand forecasting model It can be calculated by the following formula:

[0164]

[0165] Among them, W p and b p are the weight matrix and bias term of the neural network respectively, and σ is the activation function (such as ReLU or Sigmoid). is the predicted value of power demand at timestamp t. Through this model, the present invention can accurately predict future power demand based on spatiotemporal characteristics.

[0166] Furthermore, when obtaining the power demand forecast value After that, the present invention compares it with the actual power demand t The comparison is performed to calculate the error and the model is optimized based on the error.

[0167] Further, error calculation and feedback:

[0168] Assume that the actual power demand is y t The power demand is predicted to be The present invention adopts mean square error (MSE) as the error metric:

[0169]

[0170] Furthermore, in order to optimize the model in real time, the present invention introduces a dynamic learning mechanism. In this mechanism, the present invention uses the error Perform online learning and update model parameters through the adaptive gradient descent (AdaGrad) algorithm to reduce the prediction error. The parameter update formula is:

[0171]

[0172] Where η is the learning rate, is the gradient of the loss function with respect to the weights, G t is the sum of the squares of historical gradients, and ∈ is a small constant to prevent division by zero errors. Through this dynamic learning approach, the system can continuously optimize the power demand forecasting model based on real-time feedback.

[0173] Furthermore, this scheme can effectively capture the spatiotemporal variation characteristics of power demand through the spatiotemporal graph convolutional network (ST-GCN) combined with the innovative spatiotemporal dependency regularization term. Compared with traditional power demand forecasting methods, this scheme achieves more accurate demand modeling and forecasting through graph convolution and spatiotemporal dependency regularization, especially in the dynamic power marketing environment, and can adjust and optimize the forecast results in real time.

[0174] By closely combining spatiotemporal information with demand forecasting, the present invention can significantly improve the accuracy of power demand forecasting and provide more reliable decision support for power companies. This demand forecasting method based on spatiotemporal graph convolutional networks can not only serve the power industry, but also provide new technical ideas for spatiotemporal data analysis in other fields.

[0175] S3. Use the deep reinforcement learning model to perform dynamic strategy optimization based on the prediction results of the neural network model.

[0176] Specifically, in this step, the input data comes from the previous step, that is, the global spatiotemporal feature vector obtained by the spatiotemporal graph convolutional network (ST-GCN) These features contain the temporal and spatial distribution information of power demand, and can reflect the power demand and change trend of each user and time period in the power system. Based on this output, the present invention will construct a reinforcement learning (DRL) environment.

[0177] The core task of this environment is to In this case, the goal of the decision-making agent is to maximize the overall dispatching benefit, taking into account multiple factors such as the error between actual demand and predicted demand, energy dispatching cost, and system stability. Specifically, the spatiotemporal feature vector It will be used as the state space input for reinforcement learning.

[0178] State Space: The state space contains the global spatiotemporal characteristics at each time t These features combine information about spatial distribution (user location and demand) and temporal variation (historical demand trends). Specifically, is a vector of dimension d, where d represents the dimension of spatiotemporal features.

[0179] Action Space: The action space defines the actions that the agent can perform at each time step. For example, an action can represent adjusting the parameters of power dispatch or pricing strategy. Each action a t It can be a discrete value (such as adjusting the electricity price) or a continuous value (such as adjusting the supply). In this scheme, the action space is assumed to be a set of discrete actions, representing different scheduling strategies.

[0180] Reward Function: The reward function is a key part in reinforcement learning. The goal of this invention is to allow the agent to obtain feedback based on the impact of its decision on power demand scheduling. In this solution, the reward function will consist of the following two parts:

[0181] Demand error term: This is the error between predicted demand and actual demand, which encourages the model to make more accurate demand forecasts.

[0182] Energy cost item: This is the cost of electricity resources consumed by executing the scheduling strategy, ensuring that the intelligent agent reduces resource waste while optimizing demand.

[0183] Therefore, the reward function r t It can be expressed as:

[0184]

[0185] in, is the predicted electricity demand, y t It is a practical need. Is to perform an action , λ1 and λ2 are the weight coefficients of control error and cost respectively.

[0186] Furthermore, in this step, the present invention uses a deep Q-network (DQN) to design a deep reinforcement learning model to optimize the power dispatch strategy. DQN is a reinforcement learning method that approximates the Q-value function through a deep neural network. To evaluate the status Next, take action a t After that, the expected future reward.

[0187] Q value update rule: The core of DQN is the Q value update formula. By iteratively learning the Q value function, the agent can find the optimal scheduling strategy. The Q value update formula is as follows:

[0188]

[0189] Among them, α is the learning rate, γ is the discount factor, and rt It's a reward. The next state By continuously updating the Q value, the agent can gradually learn the optimal power dispatching strategy through historical experience.

[0190] Furthermore, in order to approximate the Q-value function, the present invention uses a model consisting of two layers of fully connected neural networks (FCN). The network structure is as follows:

[0191]

[0192] Among them, NN is a neural network, and W represents the weight in the network. The input of the neural network is the spatiotemporal features The output is the Q value of different actions. In order to further improve the expressiveness of the Q value function, more complex network structures such as convolutional neural networks (CNN) or graph convolutional networks (GCN) can be considered to better capture the correlation between spatiotemporal features.

[0193] Furthermore, in order to better simulate the spatiotemporal characteristics of power demand and avoid overfitting, the present invention introduces a spatiotemporal dependency regularization term. This regularization term can effectively constrain power dispatching and prevent the agent from making incoherent dispatching decisions, thereby ensuring that the dispatching strategy conforms to the spatiotemporal distribution characteristics of the power system.

[0194] Furthermore, suppose the adjacency matrix of user i and user j in the space-time graph is A ij , this regularization term improves the spatiotemporal consistency of decisions by encouraging the scheduling strategy differences between adjacent users to be minimized. It can be expressed as:

[0195]

[0196] Among them, λ3 is the regularization hyperparameter, A ij is the similarity between user i and user j (from the adjacency matrix of the spatiotemporal graph), and They represent the scheduling actions of user i and user j at time step t respectively.

[0197] The introduction of this regularization term enables the model to not only optimize the demand forecast error and cost, but also ensure the spatial consistency of scheduling decisions through spatiotemporal constraints, avoiding unnecessary waste of resources due to excessive adjustment of scheduling strategies for certain time periods or certain users.

[0198] Furthermore, in order to ensure the stability of the reinforcement learning model and avoid the oscillation of the Q value, the present invention adopts the target network (TargetNetwork) and experience replay (ExperienceReplay) technology.

[0199] Target network: Target network is a common technique in reinforcement learning. It slows down the Q value update speed by fixing the parameters of the target network, thereby avoiding instability during training. The Q value update rule of the target network is as follows:

[0200]

[0201] Where Q′ represents the target network and Q represents the current network. The weight of the target network is updated at fixed intervals to reduce the oscillation of the Q value.

[0202] Experience replay: In the training process of reinforcement learning, experience replay can effectively increase the diversity of samples and avoid overfitting of the agent. After each interaction with the environment, the agent stores the current state, action, reward, and next state in the experience pool. By randomly extracting small batches of samples from the experience pool for training, the agent can learn in a wider experience space, thereby improving the generalization ability of the strategy.

[0203] Furthermore, in order to further improve the real-time and accuracy of power dispatch, the present invention designs a feedback mechanism to dynamically adjust the power dispatch strategy by monitoring the difference between power demand and predicted demand in real time. After each decision, the agent fine-tunes the strategy according to the error between actual demand and predicted demand.

[0204] Feedback update mechanism: Assume that at time step t, the scheduling decision chosen by the agent is a t , the actual power demand is y t , the forecast demand is The present invention designs a feedback-based strategy update formula:

[0205]

[0206] Where ΔW is the weight update, η is the learning rate, is the gradient of the Q value, and t The feedback mechanism enables the model to adjust its strategy according to real-time data after each scheduling, thereby further improving the accuracy of prediction and scheduling.

[0207] Furthermore, this scheme combines the spatiotemporal graph convolutional network (ST-GCN) and deep reinforcement learning (DQN) technology, and innovatively uses the spatiotemporal dependency regularization term to optimize the power dispatch strategy, thereby improving the utilization efficiency of power resources and reducing the dispatch error. Through the combination of feedback mechanism and target network, the model can adapt to dynamically changing power demand and quickly adjust the strategy. In the context of power demand forecasting and resource scheduling, this scheme provides a new intelligent optimization method with strong application potential.

[0208] The present invention introduces a spatiotemporal dependency regularization term so that the dispatching decision not only considers the error of power demand, but also the spatiotemporal correlation. The optimal strategy learning of power dispatching is realized through the reinforcement learning model (DQN), avoiding the limitations of traditional methods. The power dispatching strategy is dynamically adjusted using a feedback mechanism to achieve real-time optimization. The target network and experience replay technology are introduced to ensure the stability and generalization ability of model training. These innovations provide a more efficient and flexible optimization solution for power dispatching and demand management, which can cope with complex demand changes and large-scale system applications.

[0209] S4. Collect feedback data after execution strategy optimization in real time, design a strategy evaluation function to evaluate the effectiveness of the marketing strategy, and modify the deep reinforcement learning model according to the evaluation results; the feedback data includes actual power demand, system response status and predicted power demand.

[0210] Specifically, the input comes from the output a of the deep reinforcement learning (DRL) dynamic decision optimization in the previous stage t , that is, the agent is based on the global spatiotemporal features at time step t Power dispatch decisions made.

[0211] Real-time feedback data includes:

[0212] Actual power demand t ∈R N , represents the actual power demand of all N users at time step t.

[0213] System response status t ∈R M , represents the system state at time step t (such as load fluctuation, power generation capacity, etc.).

[0214] Forecasting electricity demand The electricity demand predicted by the demand forecasting model in the previous step.

[0215] Furthermore, the scheduling error E is calculated t and system stability S t , used to evaluate the effect of the current scheduling strategy.

[0216] Among them, the scheduling error E t The calculation is as follows:

[0217]

[0218] Among them, y t and are the actual and predicted electricity demand, respectively.

[0219] System stability t The calculation is as follows:

[0220]

[0221] in, is the i-th dimension of the system state, and M is the system state dimension.

[0222] Furthermore, the reward function R is designed based on the scheduling error and system stability. t , used to give feedback to the agent:

[0223] R t =-λ1E t -λ2S t

[0224] Among them, λ1,λ2 are the weighted coefficients of scheduling error and system stability in the reward function.

[0225] Dynamic policy adjustment and retraining

[0226] Use the TemporalDifference (TD) method to update the Q value function:

[0227]

[0228] Among them, α is the learning rate, R t It's a reward.

[0229] Furthermore, in order to enhance the stability of the system, a nonlinear regularization term is introduced, and the modified Q value update formula is as follows:

[0230]

[0231] Among them, β is the regularization coefficient, S t is a measure of system stability.

[0232] Furthermore, based on the current strategy evaluation function P t , quantify the overall effect of the current scheduling strategy:

[0233]

[0234] Among them, α1, α2, and α3 are hyperparameters used to balance scheduling error, system stability, and total scheduling. is the power allocation of the agent to the i-th user.

[0235] Furthermore, training auxiliary models Used to predict the long-term response of the system under different scheduling strategies:

[0236]

[0237] in, It is an auxiliary model obtained through training, which is used to predict the impact of scheduling decisions on the long-term stability of the system.

[0238] Furthermore, the system performance is continuously monitored, and the power dispatch strategy is dynamically adjusted in combination with real-time feedback and evaluation indicators to ensure a balance between system stability and dispatch accuracy over multiple time steps.

[0239] Adjust scheduling strategies based on actual feedback to reduce errors and optimize system stability.

[0240] Through the above steps, the scheme can achieve adaptive optimization in power dispatching, reduce dispatching errors, ensure system stability, and adjust strategies based on real-time feedback to ensure that the intelligent agent continuously optimizes dispatching behavior in complex environments.

[0241] S5. Based on real-time feedback data, the optimized strategy is applied to different users according to their personalized needs to ensure that each user receives the recommendation that best meets their needs.

[0242] Specifically, in this stage, the input is the scheduling strategy optimization result a from the previous stage t , System Status t and real-time feedback data t Through these data, personalized needs can be identified and modeled for each user or user group, further refining user needs.

[0243] User demand modeling: Use historical power load, response mode, preference and other data combined with time series analysis to build a personalized demand vector This demand vector includes the user's electricity consumption behavior characteristics and the user's demand for dispatch response:

[0244]

[0245] in, is the power load of user i at time t. is the time delay for the user to respond to the scheduling request. is the maximum response capability of the user at time t.

[0246] is the user's preferred scheduling period, representing the time window that he or she prioritizes.

[0247] Furthermore, the user groups are divided into different demand groups through clustering algorithms (such as K-means or density-based clustering algorithm DBSCAN), which can help to perform personalized scheduling more accurately.

[0248] Furthermore, in the personalized recommendation strategy model, the core goal is to and the system global state s t , output the user's personalized scheduling strategy Thereby achieving dynamic optimization of the power system. Here, the present invention adopts a recommendation system model based on a deep neural network θ Optimize decision making.

[0249] The recommendation strategy model generates personalized scheduling decisions based on the user demand vector and system status through a deep learning model:

[0250]

[0251] Among them, f θ It is a deep neural network model obtained through training, and θ represents the weight of the network. It is a personalized dispatching strategy, which represents the power dispatching instruction that user i needs to execute at time t.

[0252] Furthermore, the user utility function It is used to quantify the user's satisfaction under a specific scheduling decision. This utility function not only considers the satisfaction of the user's power demand, but also considers the efficiency of the response and the saving of related costs. Therefore, the present invention designs a comprehensive utility function including multiple aspects:

[0253]

[0254] in, It is the load satisfaction of user i at time t, which is usually measured by the difference between the actual power supply and the user's demand. It is the response efficiency, which takes into account how quickly the user responds to the dispatch within the specified time. It is the electricity cost saved by the user, which measures the impact of the dispatch strategy on the user's economic benefits. The weight coefficients γ1, γ2, and γ3 in the user utility function reflect the relative importance of different factors to the user and depend on the user's personalized needs.

[0255] Furthermore, in order to respond to real-time changes in the system, the present invention designs a dynamic adjustment mechanism. tor user feedback t Personalized recommendation strategy when changes occur Will be updated based on real-time feedback.

[0256] Among them, after each strategy is executed, the effect of the current strategy is evaluated by calculating the real-time reward:

[0257]

[0258] in, and It is also the user load satisfaction and cost savings mentioned before; α1, α2 are adjustment factors that control the priority of different utilities.

[0259] Furthermore, based on the current rewards The personalized recommendation model f is trained by reinforcement learning or incremental training method. θ Make adjustments and update the model parameters θ to better adapt to the real-time feedback.

[0260] Furthermore, while realizing personalized dispatch, it is necessary to combine personalized strategy with global dispatch strategy to ensure the coordination and optimization of the entire power system. Therefore, the present invention designs a weighted fusion mechanism to combine the global strategy a t Personalization strategy Perform synthesis to generate the final scheduling strategy By introducing weight coefficients λ1 and λ2, the global strategy and personalized strategy are weighted and integrated according to different priorities:

[0261]

[0262] Among them, a t It is a global dispatch strategy to ensure that the power system is optimally dispatched at the global level; It is a personalized scheduling strategy that is optimized for the needs of specific users or user groups.

[0263] Furthermore, the weighting coefficients λ1 and λ2 can be dynamically adjusted according to system load, changes in user demand and real-time feedback to ensure the flexibility and adaptability of the system.

[0264] Furthermore, the long-term personalized scheduling strategy needs to be adjusted and optimized through periodic effect evaluation. At the end of each cycle, the present invention evaluates the long-term benefits of the personalized strategy and updates the strategy based on the continuous feedback from users. By periodically calculating the average value of the user utility function To evaluate the long-term effectiveness of the strategy:

[0265]

[0266] Where T is the length of the evaluation period, is the user utility value at each time t.

[0267] According to the long-term feedback results, the parameters θ of the personalized recommendation model are further adjusted through optimization algorithms (such as genetic algorithms or simulated annealing) to improve the overall benefit.

[0268] Furthermore, as the power dispatching system continues to evolve, the diversity and real-time nature of user needs require the system to be continuously optimized. Whenever the system encounters new load patterns, changes in user demand, or unforeseen emergencies, the personalized recommendation system needs to be adjusted in a timely manner.

[0269] Adaptive regularization terms are introduced to ensure efficient updating of the personalized recommendation model under real-time feedback:

[0270]

[0271] in, is a regularization term that controls the difference between the personalized strategy and the global strategy; β1, β2 are regularization factors that control the weights of the two regularization terms.

[0272] like Figure 2 As shown, in another embodiment of the present invention, a data monitoring system for power marketing is provided, the system comprising:

[0273] The power marketing data collection unit 101 is used to collect raw data from multiple power marketing data sources, standardize and preprocess the raw data, and fuse the data by weighted aggregation after preprocessing to obtain weighted fused features. The multimodal feature X of each timestamp t is t Integrate into a unified feature representation X to form a data set X = {X1,X2,…,X t ,X T};

[0274] A space-time graph construction unit 102 is used to construct a space-time graph, update node features of the space-time graph using a graph convolutional neural network, introduce a space-time dependency regularization term during the update process to improve the performance of the space-time graph convolutional network in power marketing prediction, and then aggregate all updated nodes to obtain a global space-time feature vector, and use a neural network model to predict power based on the global space-time feature vector;

[0275] The deep reinforcement learning unit 103 is used to use the deep reinforcement learning model to perform dynamic strategy optimization according to the prediction results of the neural network model, specifically including:

[0276] The global spatiotemporal feature vector is used as the state space input of reinforcement learning. The state space contains the global spatiotemporal features at each time t. The action space is defined as the operation performed by the agent at each time step. The reward function r t Including demand error term and energy cost term, it is expressed as follows:

[0277]

[0278] in, is the predicted electricity demand, y t It is a practical need. Is to perform an action The energy cost at this time, λ1 and λ2 are the weight coefficients of control error and cost respectively, i is the node i, and N is the total number of nodes;

[0279] Build a deep reinforcement learning model, including:

[0280] (a) Using DQN to optimize the power dispatch strategy, where the update rule of the DQN is expressed as follows:

[0281]

[0282] Among them, α is the learning rate, γ is the discount factor, and r t It's a reward. The next state The maximum Q value of all actions; is the state space, i.e. the global space-time eigenvector; a t It is the optimization result of scheduling strategy;

[0283] (b) A deep reinforcement learning model consisting of a two-layer fully connected neural network is used to approximate the Q-value function. The network structure is:

[0284]

[0285] Among them, NN is a neural network, and W represents the weight in the network. The input of the neural network is the spatiotemporal features The output is the Q value of different actions;

[0286] The time-space dependency regularization term is introduced to constrain power dispatching to prevent the intelligent agent from making incoherent dispatching decisions. At the same time, the standard network and experience replay technology are used to avoid the oscillation of Q value in the deep reinforcement learning process to ensure the stability of the reinforcement learning model.

[0287] Design a feedback mechanism to dynamically adjust the power dispatch strategy by monitoring the difference between power demand and predicted demand in real time. After each decision, the agent fine-tunes the strategy based on the error between actual demand and predicted demand.

[0288] A real-time optimization unit 104 is used to collect feedback data after executing strategy optimization in real time, design a strategy evaluation function to evaluate the effect of the marketing strategy, and modify the deep reinforcement learning model according to the evaluation result; the feedback data includes actual power demand, system response status and predicted power demand;

[0289] The personalized recommendation unit 105 is used to apply the optimized strategy to different users according to the personalized needs of different users based on the real-time feedback data, so as to ensure that each user gets the recommendation that best meets his or her needs.

[0290] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of the present invention. In practical applications, technicians in this field can select part or all of them according to actual needs to achieve the purpose of the present embodiment, and no limitation is made here.

[0291] In addition, for technical details that are not described in detail in this embodiment, reference can be made to the parameter operation method provided in any embodiment of the present invention, and will not be repeated here.

[0292] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0293] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0294] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0295] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A data monitoring method for power marketing, characterized in that: The method comprises: S1. Collect raw data from multiple power marketing data sources, standardize and preprocess the raw data, and fuse the data using weighted aggregation after preprocessing to obtain weighted fused features. t Integrate into a unified feature representation X to form a data set X = {X1,X2,…,X t ,X T }; S2. Constructing a spatiotemporal graph based on multimodal features, using a graph convolutional neural network to update node features of the spatiotemporal graph, and introducing a spatiotemporal dependency regularization term during the updating process to improve the performance of the spatiotemporal graph convolutional network in power marketing prediction, and then aggregating all updated nodes to obtain a global spatiotemporal feature vector, and using a neural network model to predict power based on the global spatiotemporal feature vector; S3. Use the deep reinforcement learning model to perform dynamic strategy optimization based on the prediction results of the neural network model, including: The global spatiotemporal feature vector is used as the state space input of reinforcement learning. The state space contains the global spatiotemporal features at each time t. The action space is defined as the operation performed by the agent at each time step. The reward function r t Including demand error term and energy cost term, it is expressed as follows: in, is the predicted electricity demand, y t It is a practical need. Is to perform an action The energy cost at this time, λ1 and λ2 are the weight coefficients of control error and cost respectively, i is the node i, and N is the total number of nodes; Build a deep reinforcement learning model, including: (a) Using DQN to optimize the power dispatch strategy, where the update rule of the DQN is expressed as follows: Among them, α is the learning rate, γ is the discount factor, and r t It's a reward. The next state The maximum Q value of all actions; is the state space, i.e. the global space-time eigenvector; a t It is the optimization result of scheduling strategy; (b) A deep reinforcement learning model consisting of a two-layer fully connected neural network is used to approximate the Q-value function. The network structure is: Among them, NN is a neural network, and W represents the weight in the network. The input of the neural network is the spatiotemporal features The output is the Q value of different actions; The time-space dependency regularization term is introduced to constrain power dispatching to prevent the intelligent agent from making incoherent dispatching decisions. At the same time, a two-layer fully connected neural network and experience replay technology are used to avoid Q-value oscillation during deep reinforcement learning and ensure the stability of the reinforcement learning model. Design a feedback mechanism to dynamically adjust the power dispatch strategy by monitoring the difference between power demand and predicted demand in real time. After each decision, the agent fine-tunes the strategy based on the error between actual demand and predicted demand. S4. Collect feedback data after execution strategy optimization in real time, design a strategy evaluation function to evaluate the effect of the marketing strategy, and modify the deep reinforcement learning model according to the evaluation results; the feedback data includes actual power demand, system response status and predicted power demand; S5. Based on real-time feedback data, the optimized strategy is applied to different users according to their personalized needs to ensure that each user receives the recommendation that best meets their needs.

2. A data monitoring method for power marketing according to claim 1, characterized in that: The standardization process is expressed as: Assume x ij represents the original data point, then the standardized data Calculated by the following formula: in, is the data point x ij The mean of is the data point x ij The standard deviation of The preprocessing is expressed as: Set a unified timestamp set t1, t2, ..., t T , and use the interpolation method to convert the data x in each data source ij and timestamp t j Corresponding to these unified timestamps, we get the interpolated data The weighted aggregation method is used to fuse the data to obtain the weighted fused features, which are expressed as: Assume that each data source The standardized data is Weighted fusion feature X t At timestamp t, it is calculated by the following formula: Among them, w i Represents the data source The weight of Is the data source Normalized data at timestamp t.

3. The data monitoring method for power marketing according to claim 1, characterized in that: The space-time graph is constructed as follows: Define an adjacency matrix A t To represent the spatiotemporal association between different power users at timestamp t, there are N power users and the adjacency matrix A t Element A ij represents the spatiotemporal similarity between user i and user j at timestamp t, which is calculated as follows: Among them, X i and X j are the feature vectors of user i and user j at timestamp t, σ s is the spatial similarity control parameter, represents the neighbor set of user i; The node features of the spatiotemporal graph are updated using a graph convolutional neural network, and the graph convolution operation is as follows: Among them, W t is the weight matrix at timestamp t, σ is the activation function, X′ i (t) is the updated feature; The spatiotemporal dependency regularization term R t , expressed as: Among them, λ is a hyperparameter that controls the regularization strength, A ij is the weight of the spatiotemporal association; The aggregation of all updated nodes is expressed as: in, is the aggregated spatiotemporal feature vector, and N is the number of nodes in the graph; The power is predicted using a neural network model according to the global spatiotemporal feature vector, which is represented by: in, is the predicted electricity demand, W p and b p are the weight matrix and bias term of the neural network, and σ is the activation function.

4. A data monitoring method for power marketing according to claim 3, characterized in that: Getting the predicted electricity demand Then compare it with the actual power demand t Compare to calculate the error, and optimize the neural network model according to the error, which is expressed as: At the same time, according to the error Perform online learning and update the model parameters of the neural network model through the adaptive gradient descent algorithm, which is expressed as: Where η is the learning rate, is the gradient of the loss function with respect to the weights, G t is the sum of squares of historical gradients, and ∈ is a small constant to prevent division by zero errors.

5. The data monitoring method for power marketing according to claim 1, characterized in that: The spatiotemporal regularization term is constructed as follows: Assume that the adjacency matrix of user i and user j in the space-time graph is A ij , the spatiotemporal regularization term The temporal and spatial consistency of decisions is improved by encouraging the scheduling strategy differences between adjacent users to be minimized, which can be expressed as: Among them, λ3 is the regularization hyperparameter, A ij is the similarity between user i and user j, and Respectively represent the scheduling actions of user i and user j at time step t; The Q value update rule of the target network is as follows: Among them, Q′ represents the target network, Q represents the current network; The experience replay is that after each interaction with the environment, the agent stores the current state, action, reward and next state into the experience pool; The feedback mechanism is designed as follows: Assume that at time step t, the scheduling decision chosen by the agent is a t , the actual power demand is y t , the forecast demand is The feedback-based strategy update is expressed as: Where ΔW is the weight update, η is the learning rate, is the gradient of the Q value, and t Forecast and actual demand respectively.

6. The data monitoring method for power marketing according to claim 1, characterized in that: The S4 specifically includes: Calculate the dispatch error based on the actual and predicted power demand, calculate the system stability based on the system response state, and design the reward function R based on the dispatch error and system stability t Feedback to the agent for dynamic strategy adjustment and retraining; wherein the scheduling error E t The calculation is as follows: The system stability S t The calculation is as follows: in, is the i-th dimension of the system state, and M is the system state dimension; The reward function R t , expressed as: R t =-λ1E t -λ2S t Among them, λ1,λ2 are the weighted coefficients of scheduling error and system stability in the reward function; Design the current policy evaluation function P t , quantify the overall effect of the current scheduling strategy, expressed as: Among them, α1, α2, and α3 are hyperparameters used to balance scheduling error, system stability, and total scheduling. is the power allocation of the agent to the i-th user; Training auxiliary models Used to predict the long-term response of the system under different scheduling strategies: in, It is an auxiliary model obtained through training, which is used to predict the impact of scheduling decisions on the long-term stability of the system.

7. A data monitoring method for power marketing according to claim 6, characterized in that: The designed reward function R t Feedback is given to the agent for dynamic strategy adjustment and retraining, specifically: Update the Q value function using the TemporalDifference method: Among them, α is the learning rate, R t It's a reward. The nonlinear regularization term is introduced to enhance the stability of the system, and the modified Q value update formula is as follows: Among them, β is the regularization coefficient.

8. The data monitoring method for power marketing according to claim 1, characterized in that: The S5 specifically includes: Collect user needs, combine time series analysis, and build personalized demand vectors And through the clustering algorithm, the user group is divided into different demand groups; the user demand includes historical power load, response mode and preference; A personalized recommendation strategy model is built using a deep neural network to generate personalized scheduling decisions based on user demand vectors and system status; Design a comprehensive utility function to evaluate personalized scheduling decisions; Whenever the system state or user feedback changes, the personalized scheduling decision is updated by calculating real-time rewards to evaluate the effectiveness of the current strategy; Combine personalized dispatch decisions with global dispatch strategies to coordinate and optimize power marketing data; At the end of each cycle, the long-term benefits of personalized scheduling decisions are evaluated and the strategy is updated based on continuous feedback from users. At the same time, whenever the global scheduling strategy encounters an emergency, the personalized scheduling decision needs to be adjusted in a timely manner.

9. A data monitoring method for power marketing according to claim 8, characterized in that: The comprehensive utility function is expressed as: in, is the load satisfaction of user i at time t; is the response efficiency, which takes into account how quickly the user responds to the dispatch within the specified time; is the electricity cost saved by the user, which measures the impact of the dispatch strategy on the user's economic benefits; γ1, γ2, and γ3 are weight coefficients, reflecting the relative importance of different factors to the user.

10. A data monitoring system for power marketing, characterized in that: The system comprises: The power marketing data acquisition unit is used to collect raw data from multiple power marketing data sources, standardize and preprocess the raw data, and fuse the data using weighted aggregation after preprocessing to obtain weighted fused features. The multimodal feature X of each timestamp t is t Integrate into a unified feature representation X to form a data set X = {X1,X2,…,X t ,X T }; A space-time graph construction unit is used to construct a space-time graph, update node features of the space-time graph using a graph convolutional neural network, introduce a space-time dependency regularization term during the update process to improve the performance of the space-time graph convolutional network in power marketing prediction, and then aggregate all updated nodes to obtain a global space-time feature vector, and use a neural network model to predict power based on the global space-time feature vector; The deep reinforcement learning unit is used to perform dynamic strategy optimization using the deep reinforcement learning model based on the prediction results of the neural network model, including: The global spatiotemporal feature vector is used as the state space input of reinforcement learning. The state space contains the global spatiotemporal features at each time t. The action space is defined as the operation performed by the agent at each time step. The reward function r t Including demand error term and energy cost term, it is expressed as follows: in, is the predicted electricity demand, y t It is a practical need. Is to perform an action The energy cost at this time, λ1 and λ2 are the weight coefficients of control error and cost respectively, i is the node i, and N is the total number of nodes; Build a deep reinforcement learning model, including: (a) Using DQN to optimize the power dispatch strategy, where the update rule of the DQN is expressed as follows: Among them, α is the learning rate, γ is the discount factor, and r t It's a reward. The next state The maximum Q value of all actions; is the state space, i.e. the global space-time eigenvector; a t It is the optimization result of scheduling strategy; (b) A deep reinforcement learning model consisting of a two-layer fully connected neural network is used to approximate the Q-value function. The network structure is: Among them, NN is a neural network, and W represents the weight in the network. The input of the neural network is the spatiotemporal features The output is the Q value of different actions; The time-space dependency regularization term is introduced to constrain power dispatching to prevent the intelligent agent from making incoherent dispatching decisions. At the same time, a two-layer fully connected neural network and experience replay technology are used to avoid Q-value oscillation during deep reinforcement learning and ensure the stability of the reinforcement learning model. Design a feedback mechanism to dynamically adjust the power dispatch strategy by monitoring the difference between power demand and predicted demand in real time. After each decision, the agent fine-tunes the strategy based on the error between actual demand and predicted demand. A real-time optimization unit is used to collect feedback data after executing strategy optimization in real time, design a strategy evaluation function to evaluate the effect of the marketing strategy, and modify the deep reinforcement learning model according to the evaluation results; the feedback data includes actual power demand, system response status and predicted power demand; The personalized recommendation unit is used to apply the optimized strategy to different users according to their personalized needs based on real-time feedback data, ensuring that each user receives the recommendation that best meets their needs.

Images