Power supply service optimization method and device for smart city

Abstract

The invention provides a smart city-oriented power supply service optimization method and device. The method comprises the steps of 1, preprocessing multi-modal data; step 2, carrying out cross-modal feature fusion; 3, dynamically predicting the power consumption; and step 4, power supply strategy optimization. According to the method, through multi-modal data fusion and dynamic user behavior modeling, the bottlenecks of high data isomerism, insufficient cross-domain collaboration, difficulty in random behavior quantification and the like in a traditional method are overcome. According to the method, multi-source heterogeneous data such as power sensor data, policy texts and meteorological information are deeply fused, a time sequence prediction model and user behavior probability analysis of a non-homogeneous Markov chain are combined, a cross-domain collaborative intelligent optimization framework is constructed, comprehensive perception and accurate prediction of power supply data in a power system link are realized, and the power supply efficiency is improved. The prediction model is used to predict and adjust the power consumption demand, the distribution and balance of the power consumption load are optimized, and the power consumption service quality and the customer satisfaction are improved.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and electronic information, and specifically to a method and device for optimizing power supply services for smart cities. Background Art

[0002] With the deepening reform of the power industry, power supply companies must transform into service-oriented enterprises, establish a customer-oriented management mechanism, and improve customer service management to achieve higher customer satisfaction and loyalty. At present, how to promote the deep integration of digital technology and public services in the power industry, draw a blueprint for digital, clean, personalized, and convenient power supply services, and enhance people's sense of gain and satisfaction with electricity services has become a difficult problem that needs to be solved in the current power supply services. At present, there is little research on multimodal data fusion prediction technology for power supply services, and the main problems are as follows:

[0003] (1) Data acquisition and processing are difficult: Multimodal data comes from a wide range of sources, including sensor data, image data, and text data. The acquisition and processing of these data requires complex technologies and equipment. Data quality and consistency are also a challenge. Data from different sources may have inconsistent formats and accuracy.

[0004] (2) Immature technology: Multimodal data fusion technology is still in its early stages of development, and many algorithms and models are still being optimized and improved. Cross-domain data fusion requires the integration of multiple technologies, such as machine learning, deep learning, and data mining. The integration and application of these technologies are not yet mature.

[0005] (3) Insufficient cross-domain cooperation: Multimodal data fusion requires cooperation in multiple fields such as electricity, policy, and meteorology, but the current cross-domain cooperation mechanism is not yet perfect. There are also certain obstacles to data sharing and standardization between different fields. Summary of the Invention

[0006] In view of this, the present invention provides a power supply service optimization method and device for smart cities. Through power load forecasting and network integration, data from different sources and in different forms and structures (such as power sensor data, policy text data, user behavior probability data, etc.) are integrated to achieve comprehensive perception and accurate prediction of power supply data in the power system links, use prediction models to predict and adjust electricity demand, optimize power load distribution and balance, and improve power service quality and customer satisfaction.

[0007] A method for optimizing power supply services for smart cities, comprising:

[0008] Step 1: Multimodal data preprocessing

[0009] Performing sliding window segmentation on the power system time series data to generate a supervised learning sample set with time continuity, and performing Z-score normalization processing on the sample set;

[0010] Perform Chinese word segmentation and stop word filtering operations on the text data, input the processed word segmentation sequence into the pre-trained Word2Vec model to generate a text vectorization matrix;

[0011] Align the user behavior data with the American Time Use Survey (ATUS) dataset for behavior labels and define a set of discrete electricity consumption behavior states.

[0012] Step 2: Cross-modal feature fusion

[0013] Extracting local periodic features from the normalized time series data through a first convolutional neural network, and outputting a time series feature tensor with a dimension of T×F;

[0014] Extracting semantic association features from the text vectorization matrix through a second convolutional neural network, and outputting a text feature tensor with a dimension of K×D;

[0015] The temporal feature tensor and the text feature tensor are concatenated at the channel level in the flattening layer to generate a joint feature vector with a dimension of (T×F+K×D);

[0016] Step 3: Dynamic prediction of electricity consumption

[0017] The joint feature vector is input into a long short-term memory network (LSTM), and the time dependency is captured through a hidden state gating mechanism to output a predicted power consumption sequence;

[0018] The root mean square error (RMSE) and mean absolute percentage error (MAPE) loss function are used to optimize the LSTM network parameters.

[0019] Step 4: Power supply strategy optimization

[0020] Based on the non-homogeneous Markov chain, the user behavior state transition probability matrix is calculated, and the Monte Carlo method is used to simulate the dynamic changes of behavior in different time periods;

[0021] Performing a tensor multiplication operation on the predicted power consumption sequence output by the LSTM and the probability transfer matrix to generate a joint decision feature vector;

[0022] The dynamic regulation potential of the user's electricity consumption behavior is quantitatively evaluated by the joint decision feature vector, and the electricity load distribution is optimized based on the evaluation result of the dynamic regulation potential of the household's electricity consumption behavior.

[0023] Furthermore, the power system time series data in step 1 includes voltage and power.

[0024] Furthermore, the text data in step one includes policy data and weather data.

[0025] Furthermore, in step 4, the dynamic adjustment potential of the user's electricity consumption behavior is quantitatively evaluated by the joint decision feature vector, specifically including:

[0026] (1) Align the peak electricity consumption period predicted by LSTM with the user behavior state transition probability matrix, analyze the user's device adjustability during the high electricity price period, and combine the time-of-use electricity price rules in the policy data with weather data to evaluate the user's demand response potential;

[0027] (2) Using the user behavior state transition probability matrix to predict the user's out-of-home period, identify the interruptible load of non-essential equipment, and calculate the theoretical load reduction range;

[0028] (3) Generate a dynamic regulation potential distribution map based on the user demand response potential and the theoretical reduction range to provide a priority basis for the demand response plan.

[0029] Furthermore, in step 4, the electricity load distribution is optimized based on the evaluation results of the dynamic regulation potential of household electricity consumption behavior, specifically including:

[0030] (1) Dynamically adjust the operating time of adjustable load equipment based on the dynamic regulation potential of user electricity consumption behavior, and combine it with distributed energy output planning to maximize the utilization of low-cost electricity resources;

[0031] (2) Based on the peak power consumption period predicted by LSTM, the grid dispatch strategy is optimized in real time to balance the supply and demand relationship;

[0032] (3) Generate multiple sets of random power consumption scenarios through Monte Carlo simulation to verify the robustness of the scheduling strategy and ensure system stability under abnormal conditions;

[0033] (4) Smart meters are used to collect user electricity consumption data in real time, form a closed-loop feedback loop with dynamic power consumption forecasts, dynamically modify scheduling instructions, and ultimately generate time-of-use electricity price strategies and equipment control plans. Personalized prompts are pushed through user terminals to achieve load balancing and improve service quality.

[0034] A power supply service optimization device for smart cities, comprising:

[0035] Multimodal data preprocessing module, used for:

[0036] Performing sliding window segmentation on the power system time series data to generate a supervised learning sample set with time continuity, and performing Z-score normalization processing on the sample set;

[0037] Perform Chinese word segmentation and stop word filtering operations on the text data, input the processed word segmentation sequence into the pre-trained Word2Vec model to generate a text vectorization matrix;

[0038] Align the user behavior data with the American Time Use Survey (ATUS) dataset for behavior labels and define a set of discrete electricity consumption behavior states.

[0039] Cross-modal feature fusion module, used for:

[0040] Extracting local periodic features from the normalized time series data through a first convolutional neural network, and outputting a time series feature tensor with a dimension of T×F;

[0041] Extracting semantic association features from the text vectorization matrix through a second convolutional neural network, and outputting a text feature tensor with a dimension of K×D;

[0042] The temporal feature tensor and the text feature tensor are concatenated at the channel level in the flattening layer to generate a joint feature vector with a dimension of (T×F+K×D);

[0043] The power consumption dynamic prediction module is used to:

[0044] The joint feature vector is input into a long short-term memory network (LSTM), and the time dependency is captured through a hidden state gating mechanism to output a predicted power consumption sequence;

[0045] The root mean square error (RMSE) and mean absolute percentage error (MAPE) loss function are used to optimize the LSTM network parameters.

[0046] Power supply strategy optimization module, used for:

[0047] Based on the non-homogeneous Markov chain, the user behavior state transition probability matrix is calculated, and the Monte Carlo method is used to simulate the dynamic changes of behavior in different time periods;

[0048] Performing a tensor multiplication operation on the predicted power consumption sequence output by the LSTM and the probability transfer matrix to generate a joint decision feature vector;

[0049] The dynamic regulation potential of the user's electricity consumption behavior is quantitatively evaluated by the joint decision feature vector, and the electricity load distribution is optimized based on the evaluation result of the dynamic regulation potential of the household's electricity consumption behavior.

[0050] Furthermore, the power system time series data includes voltage and power.

[0051] Furthermore, the text data includes policy data and weather data.

[0052] Furthermore, the power supply strategy optimization module quantitatively evaluates the dynamic adjustment potential of the user's power consumption behavior through the joint decision feature vector, specifically including:

[0053] (1) Align the peak electricity consumption period predicted by LSTM with the user behavior state transition probability matrix, analyze the user's device adjustability during the high electricity price period, and combine the time-of-use electricity price rules in the policy data with weather data to evaluate the user's demand response potential;

[0054] (2) Using the user behavior state transition probability matrix to predict the user's out-of-home period, identify the interruptible load of non-essential equipment, and calculate the theoretical load reduction range;

[0055] (3) Generate a dynamic regulation potential distribution map based on the user demand response potential and the theoretical reduction range to provide a priority basis for the demand response plan.

[0056] Furthermore, the power supply strategy optimization module optimizes the power load distribution based on the evaluation results of the dynamic adjustment potential of household power consumption behavior, specifically including:

[0057] (1) Dynamically adjust the operating time of adjustable load equipment based on the dynamic regulation potential of user electricity consumption behavior, and combine it with distributed energy output planning to maximize the utilization of low-cost electricity resources;

[0058] (2) Based on the peak power consumption period predicted by LSTM, the grid dispatch strategy is optimized in real time to balance the supply and demand relationship;

[0059] (3) Generate multiple sets of random power consumption scenarios through Monte Carlo simulation to verify the robustness of the scheduling strategy and ensure system stability under abnormal conditions;

[0060] (4) Smart meters are used to collect user electricity consumption data in real time, form a closed-loop feedback loop with dynamic power consumption forecasts, dynamically modify scheduling instructions, and ultimately generate time-of-use electricity price strategies and equipment control plans. Personalized prompts are pushed through user terminals to achieve load balancing and improve service quality.

[0061] This invention overcomes bottlenecks in traditional methods, such as high data heterogeneity, insufficient cross-domain collaboration, and difficulty quantifying random behavior, through multimodal data fusion and dynamic user behavior modeling. By deeply integrating heterogeneous data from multiple sources, including power sensor data, policy texts, and meteorological information, and combining time series prediction models with user behavior probability analysis using non-homogeneous Markov chains, this invention constructs a cross-domain collaborative intelligent optimization framework, enabling comprehensive perception and accurate prediction of power supply data within the power system. By predicting and regulating power demand, this approach optimizes power load distribution and balance, improving power service quality and customer satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1This is a flow chart of a method for optimizing power supply services for smart cities according to an embodiment of the present invention;

[0063] Figure 2 Schematic diagram of text data feature extraction according to an embodiment of the present invention;

[0064] Figure 3 This is a schematic diagram of feature fusion according to an embodiment of the present invention;

[0065] Figure 4 This is a predictive modeling framework based on multimodal information fusion in an embodiment of the present invention.

[0066] Figure 5 Schematic diagram of probability distribution of different activities according to an embodiment of the present invention. DETAILED DESCRIPTION

[0067] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are 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 making creative efforts shall fall within the scope of protection of the present invention.

[0068] See also Figure 1 The embodiment of the present invention provides a method for optimizing power supply services for smart cities, comprising the following steps:

[0069] Step 1: Multimodal data preprocessing. Step 1 specifically includes:

[0070] Step 1.1, perform sliding window segmentation on the power system time series data (such as voltage, power), generate a supervised learning sample set with time continuity, and perform Z-score normalization on the sample set; specifically, first fill in the missing values in the original data to ensure the integrity of the data; secondly, on this basis, in order to better meet the data structure of the convolutional neural network input layer, we use sliding window technology to further process the data format. Through this method, the time series data is converted into one-dimensional data, and the solution of the time series problem is converted into the solution of the supervision problem. Finally, in order to map the data within a fixed range, the present invention uses the maximum and minimum normalization method to process the input data, and the maximum and minimum formula is shown in formula (1).

[0071]

[0072] Step 1.2: Perform Chinese word segmentation and stop word filtering on the policy data and weather data, input the processed word segmentation sequence into the pre-trained Word2Vec model to generate a text vectorization matrix;

[0073] Specifically, the original text data (policy data and weather data) is unstructured data. Before feature extraction, it needs to be processed into data. When converting the text into data, it is also necessary to retain the effective information as much as possible. Basic text data preprocessing (such as Figure 2 The process (shown in Figure 2) includes steps such as word segmentation and stop word removal. The Word2Vec model is then used to represent the text, known as word embedding. This allows for quantitative measurement of relationships between words and for discovering connections between them. This text vectorization eliminates the tedious work of general feature processing and maximizes the preservation of the original text information.

[0074] Step 1.3: Align the user behavior data with the American Time Use Survey (ATUS) dataset for behavior labels and define a discretized set of electricity usage behavior states. Specifically, the American Time Use Survey (ATUS) is used as the source of user historical data, and daily activities are divided into different electricity usage behaviors.

[0075] ATUS is a national survey sponsored by the U.S. Bureau of Labor Statistics and conducted annually by the U.S. Census Bureau. It aims to understand how U.S. residents allocate their time daily (from 4:00 AM to 4:00 AM the following morning). The survey records how residents spend time on various activities, including sleep, work, leisure, housework, and socializing. This paper uses ATUS2015, a data set consisting of 10,886 respondents and 212,605 activity records, to build a resident behavior simulation model. Because ATUS was not developed specifically for assessing residential energy use, much of the defined activity data is not relevant to smart device usage. To simplify analysis and modeling, this paper redefined 10 typical user activity states, as shown in Table 1, and distributed these states across 1,440 time periods, meaning that users are in one of these states every minute. The sleep and non-power-using behavior states can be used to define activity states where a resident is at home but does not use any specific power devices. The final activity category, away from home, refers to when the user is away from home.

[0076] Table 1 Typical activities and corresponding equipment

[0077]

[0078] These states may result in device usage or energy consumption. Seven activities were selected as representative behaviors: washing, laundry, cleaning, cooking, dishwashing, watching TV, and using a computer, as they are associated with the most common energy consumption in residential settings. The corresponding device usage behaviors and typical power values are shown in Table 2. In addition to the smart devices listed in the table, ambient smart devices such as lighting, air conditioners, and humidifiers should also be considered during residents' stays at home.

[0079] Table 2 Household appliances and corresponding power loads

[0080]

[0081] Step 2: Cross-modal feature fusion. Step 2 specifically includes:

[0082] Step 2.1. Extract local periodic features from the normalized time series data through the first convolutional neural network (CNN), and output a time series feature tensor with a dimension of T×F; use CNN to extract the implicit features of the processed time series data, using a traditional network structure, including input layer, convolution layer, pooling layer, fully connected layer and output layer from front to back.

[0083] Step 2.2: Extract semantic association features from the text vectorization matrix through a second CNN, and output a text feature tensor with a dimension of K×D; extract text word vectors from the text data through word vector training, and extract implicit features of the text data through the application of CNN in the field of natural language.

[0084] Step 2.3, the time series feature tensor and the text feature tensor are concatenated at the channel level in the flattening layer (e.g. Figure 3 As shown in Figure 2, a joint feature vector of dimension (T × F + K × D) is generated. The flattened layer is then connected to extract the features of the two data and fused, which serves as the input data for the prediction module. Due to the heterogeneity of the original data used in this invention, a direct concatenation vector fusion method is chosen to fuse the implicit features of the two data.

[0085] Step 3: Dynamic forecast of electricity consumption. Step 3 specifically includes (such as Figure 4 shown):

[0086] Step 3.1: Input the joint feature vector into a long short-term memory network (LSTM), capture the time dependency through a hidden state gating mechanism, and output a predicted power consumption sequence;

[0087] Step 3.2: Optimize the LSTM network parameters using the root mean square error (RMSE) and mean absolute percentage error (MAPE) joint loss function.

[0088] Step 4: Power supply strategy optimization. Step 4 specifically includes:

[0089] Step 4.1: Calculate the user behavior state transition probability matrix based on the non-homogeneous Markov chain and use the Monte Carlo method to simulate the dynamic changes of behavior in different time periods;

[0090] Specifically, we construct a behavior chain and use the preprocessed ATUS2015 data to construct a Markov chain of residents' electricity consumption behavior. For a random variable sequence X, we have the following definition:

[0091] X={X0,X1,X2,...,X t ,...} (2)

[0092] Where, X t Represents the random variable sequence at the current time t. Each random variable sequence X t The value set of is the same and together constitutes the state space, denoted as S. For all time instants with t ≥ 0 and all states i0, i1, i2, ..., j∈ S, the random variable sequence X has the characteristics shown in formula (3), which is called a Markov chain.

[0093] P[X t+1 =j|X t =i,X t-1 =i t-1 ,...,X0=i0]

[0094] =P[X t+1 =j|X t =i] = P i,j (3)

[0095] Where, P i,j is the probability of transitioning from state i to state j. According to the Markov chain's no-aftereffect property, when the behavior leaves state i, it must move to any state in the state space S. Therefore, each row of the matrix P must sum to 1. For each state i, the following constraints apply:

[0096]

[0097] This matrix is also called a single-step transition matrix. If the change in residential electricity consumption behavior is time-independent, then only this matrix is needed to represent the change in residential electricity consumption behavior. This matrix is called a homogeneous Markov chain. However, since the probability of users performing each activity at different times is different (e.g. Figure 5 Therefore, a time-varying Markov chain or an inhomogeneous Markov chain should be developed to flexibly simulate the dynamic electricity consumption behavior of household users.

[0098] Based on the Markov chain at different times of the day, the electricity consumption behavior of residents can be predicted by the Monte Carlo method. The Monte Carlo simulation method, also known as random simulation or statistical test method, is characterized by its ability to simulate real physical processes, convert practical problems into a probability model, and obtain approximate solutions to the problem through statistical simulation or random sampling by computer, and obtain better fitting results. In order to accurately simulate the behavior of users, the present invention uses a general random variable Y i t To describe the matrix The electricity consumption behavior state of the i-th row of the matrix It can be calculated by the following formula:

[0099]

[0100] After completing the initialization work, advance to the next moment, obtain the electricity consumption behavior state at the next moment based on the state transfer matrix of the Markov chain established previously, and then advance to the next moment, and so on, until the simulation behavior data of all users is completed.

[0101] Step 4.2: Perform a tensor multiplication operation on the predicted power consumption sequence output by the LSTM and the probability transfer matrix to generate a joint decision feature vector;

[0102] Step 4.3: quantitatively evaluate the dynamic regulation potential of the user's electricity consumption behavior through the joint decision feature vector, and optimize the electricity load distribution based on the evaluation result of the dynamic regulation potential of the household's electricity consumption behavior.

[0103] The dynamic adjustment potential of the user's electricity consumption behavior is quantitatively evaluated through the joint decision feature vector, specifically including:

[0104] (1) Align the peak electricity consumption period predicted by LSTM with the user behavior state transition probability matrix, analyze the user's equipment adjustability during high electricity price periods (such as the start-stop flexibility of air conditioners and charging piles), and combine the time-of-use electricity price rules in policy data with weather data (such as the influence coefficient of temperature on air conditioning load) to evaluate the user's demand response potential;

[0105] (2) Using the user behavior state transition probability matrix to predict the user's out-of-home period, identify the interruptible load of non-essential equipment, and calculate the theoretical load reduction range;

[0106] (3) Generate a dynamic regulation potential distribution map based on the user demand response potential and the theoretical reduction range to provide a priority basis for the demand response plan.

[0107] The optimization of electricity load distribution is based on the evaluation results of the dynamic regulation potential of household electricity consumption behavior, including:

[0108] (1) Dynamically adjust the operating time of adjustable load equipment based on the dynamic regulation potential of user electricity consumption behavior (such as washing machines and dishwashers automatically starting during low electricity price periods), combined with distributed energy (photovoltaic, energy storage) output planning to maximize the use of low-priced electricity resources;

[0109] (2) Based on the peak power consumption period predicted by LSTM, the grid dispatch strategy (such as energy storage charging and discharging sequence) is optimized in real time to balance the supply and demand relationship;

[0110] (3) Generate multiple sets of random electricity consumption scenarios (covering extreme weather and policy changes) through Monte Carlo simulation to verify the robustness of the scheduling strategy and ensure system stability under abnormal conditions;

[0111] (4) Smart meters are used to collect user electricity consumption data in real time, which is then combined with dynamic power consumption forecasts to form a closed-loop feedback loop. The dispatch instructions are then dynamically modified to ultimately generate a time-of-use electricity price strategy and equipment control plan. Personalized prompts (such as "it is recommended to turn off the air conditioner from 10:00 to 12:00") are pushed through user terminals to achieve load balancing and improve service quality.

[0112] An embodiment of the present invention further provides a power supply service optimization device for a smart city, comprising:

[0113] Multimodal data preprocessing module, used for:

[0114] Performing sliding window segmentation on the power system time series data to generate a supervised learning sample set with time continuity, and performing Z-score normalization processing on the sample set;

[0115] Perform Chinese word segmentation and stop word filtering operations on the text data, input the processed word segmentation sequence into the pre-trained Word2Vec model to generate a text vectorization matrix;

[0116] Align the user behavior data with the American Time Use Survey (ATUS) dataset for behavior labels and define a set of discrete electricity consumption behavior states.

[0117] Cross-modal feature fusion module, used for:

[0118] Extracting local periodic features from the normalized time series data through a first convolutional neural network, and outputting a time series feature tensor with a dimension of T×F;

[0119] Extracting semantic association features from the text vectorization matrix through a second convolutional neural network, and outputting a text feature tensor with a dimension of K×D;

[0120] The temporal feature tensor and the text feature tensor are concatenated at the channel level in the flattening layer to generate a joint feature vector with a dimension of (T×F+K×D);

[0121] The power consumption dynamic prediction module is used to:

[0122] The joint feature vector is input into a long short-term memory network (LSTM), and the time dependency is captured through a hidden state gating mechanism to output a predicted power consumption sequence;

[0123] The root mean square error (RMSE) and mean absolute percentage error (MAPE) loss function are used to optimize the LSTM network parameters.

[0124] Power supply strategy optimization module, used for:

[0125] Based on the non-homogeneous Markov chain, the user behavior state transition probability matrix is calculated, and the Monte Carlo method is used to simulate the dynamic changes of behavior in different time periods;

[0126] Performing a tensor multiplication operation on the predicted power consumption sequence output by the LSTM and the probability transfer matrix to generate a joint decision feature vector;

[0127] The dynamic regulation potential of the user's electricity consumption behavior is quantitatively evaluated by the joint decision feature vector, and the electricity load distribution is optimized based on the evaluation result of the dynamic regulation potential of the household's electricity consumption behavior.

[0128] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A power supply service optimization method for smart cities, characterized in that: include: Step 1: Multimodal data preprocessing Performing sliding window segmentation on the power system time series data to generate a supervised learning sample set with time continuity, and performing Z-score normalization processing on the sample set; Perform Chinese word segmentation and stop word filtering operations on the text data, input the processed word segmentation sequence into the pre-trained Word2Vec model to generate a text vectorization matrix; Align the user behavior data with the American Time Use Survey (ATUS) dataset for behavior labels and define a set of discrete electricity consumption behavior states. Step 2: Cross-modal feature fusion Extracting local periodic features from the normalized time series data through a first convolutional neural network, and outputting a time series feature tensor with a dimension of T×F; Extracting semantic association features from the text vectorization matrix through a second convolutional neural network, and outputting a text feature tensor with a dimension of K×D; The temporal feature tensor and the text feature tensor are concatenated at the channel level in the flattening layer to generate a joint feature vector with a dimension of (T×F+K×D); Step 3: Dynamic prediction of electricity consumption The joint feature vector is input into a long short-term memory network (LSTM), and the time dependency is captured through a hidden state gating mechanism to output a predicted power consumption sequence; The root mean square error (RMSE) and mean absolute percentage error (MAPE) loss function are used to optimize the LSTM network parameters. Step 4: Power supply strategy optimization Based on the non-homogeneous Markov chain, the user behavior state transition probability matrix is calculated, and the Monte Carlo method is used to simulate the dynamic changes of behavior in different time periods; Performing a tensor multiplication operation on the predicted power consumption sequence output by the LSTM and the probability transfer matrix to generate a joint decision feature vector; The dynamic regulation potential of the user's electricity consumption behavior is quantitatively evaluated by the joint decision feature vector, and the electricity load distribution is optimized based on the evaluation result of the dynamic regulation potential of the household's electricity consumption behavior.

2. The power supply service optimization method for smart cities according to claim 1, characterized in that: The power system time series data in step 1 includes voltage and power.

3. The power supply service optimization method for smart cities according to claim 1, characterized in that: The text data in step 1 includes policy data and weather data.

4. The power supply service optimization method for smart cities according to claim 1, characterized in that: In step 4, the dynamic adjustment potential of the user's electricity consumption behavior is quantitatively evaluated through the joint decision feature vector, which specifically includes: (1) Align the peak electricity consumption period predicted by LSTM with the user behavior state transition probability matrix, analyze the user's device adjustability during the high electricity price period, and combine the time-of-use electricity price rules in the policy data with weather data to evaluate the user's demand response potential; (2) Using the user behavior state transition probability matrix to predict the user's out-of-home period, identify the interruptible load of non-essential equipment, and calculate the theoretical load reduction range; (3) Generate a dynamic regulation potential distribution map based on the user demand response potential and the theoretical reduction range to provide a priority basis for the demand response plan.

5. The power supply service optimization method for smart cities according to claim 1, characterized in that: In step 4, the load distribution is optimized based on the evaluation results of the dynamic regulation potential of household electricity consumption behavior, which specifically includes: (1) Dynamically adjust the operating time of adjustable load equipment based on the dynamic regulation potential of user electricity consumption behavior, and combine it with distributed energy output planning to maximize the utilization of low-cost electricity resources; (2) Based on the peak power consumption period predicted by LSTM, the grid dispatch strategy is optimized in real time to balance the supply and demand relationship; (3) Generate multiple sets of random power consumption scenarios through Monte Carlo simulation to verify the robustness of the scheduling strategy and ensure system stability under abnormal conditions; (4) Smart meters are used to collect user electricity consumption data in real time, form a closed-loop feedback loop with dynamic power consumption forecasts, dynamically modify scheduling instructions, and ultimately generate time-of-use electricity price strategies and equipment control plans. Personalized prompts are pushed through user terminals to achieve load balancing and improve service quality.

6. A power supply service optimization device for smart cities, characterized in that: include: Multimodal data preprocessing module, used for: Performing sliding window segmentation on the power system time series data to generate a supervised learning sample set with time continuity, and performing Z-score normalization processing on the sample set; Perform Chinese word segmentation and stop word filtering operations on the text data, input the processed word segmentation sequence into the pre-trained Word2Vec model to generate a text vectorization matrix; Align the user behavior data with the American Time Use Survey (ATUS) dataset for behavior labels and define a set of discrete electricity consumption behavior states. Cross-modal feature fusion module, used for: Extracting local periodic features from the normalized time series data through a first convolutional neural network, and outputting a time series feature tensor with a dimension of T×F; Extracting semantic association features from the text vectorization matrix through a second convolutional neural network, and outputting a text feature tensor with a dimension of K×D; The temporal feature tensor and the text feature tensor are concatenated at the channel level in the flattening layer to generate a joint feature vector with a dimension of (T×F+K×D); The power consumption dynamic prediction module is used to: The joint feature vector is input into a long short-term memory network (LSTM), and the time dependency is captured through a hidden state gating mechanism to output a predicted power consumption sequence; The root mean square error (RMSE) and mean absolute percentage error (MAPE) loss function are used to optimize the LSTM network parameters. Power supply strategy optimization module, used for: Based on the non-homogeneous Markov chain, the user behavior state transition probability matrix is calculated, and the Monte Carlo method is used to simulate the dynamic changes of behavior in different time periods; Performing a tensor multiplication operation on the predicted power consumption sequence output by the LSTM and the probability transfer matrix to generate a joint decision feature vector; The dynamic regulation potential of the user's electricity consumption behavior is quantitatively evaluated by the joint decision feature vector, and the electricity load distribution is optimized based on the evaluation result of the dynamic regulation potential of the household's electricity consumption behavior.

7. The power supply service optimization device for smart cities according to claim 6, characterized in that: The power system time series data includes voltage and power.

8. The power supply service optimization device for smart cities according to claim 6, characterized in that: The text data includes policy data and weather data.

9. The power supply service optimization device for smart cities according to claim 6, characterized in that: The power supply strategy optimization module quantitatively evaluates the dynamic adjustment potential of the user's power consumption behavior through the joint decision feature vector, specifically including: (1) Align the peak electricity consumption period predicted by LSTM with the user behavior state transition probability matrix, analyze the user's device adjustability during the high electricity price period, and combine the time-of-use electricity price rules in the policy data with weather data to evaluate the user's demand response potential; (2) Using the user behavior state transition probability matrix to predict the user's out-of-home period, identify the interruptible load of non-essential equipment, and calculate the theoretical load reduction range; (3) Generate a dynamic regulation potential distribution map based on the user demand response potential and the theoretical reduction range to provide a priority basis for the demand response plan.

10. The power supply service optimization device for smart cities according to claim 6, characterized in that: The power supply strategy optimization module optimizes the power load distribution based on the evaluation results of the dynamic adjustment potential of household power consumption behavior, specifically including: (1) Dynamically adjust the operating time of adjustable load equipment based on the dynamic regulation potential of user electricity consumption behavior, and combine it with distributed energy output planning to maximize the utilization of low-cost electricity resources; (2) Based on the peak power consumption period predicted by LSTM, the grid dispatch strategy is optimized in real time to balance the supply and demand relationship; (3) Generate multiple sets of random power consumption scenarios through Monte Carlo simulation to verify the robustness of the scheduling strategy and ensure system stability under abnormal conditions; (4) Smart meters are used to collect user electricity consumption data in real time, form a closed-loop feedback loop with dynamic power consumption forecasts, dynamically modify scheduling instructions, and ultimately generate time-of-use electricity price strategies and equipment control plans. Personalized prompts are pushed through user terminals to achieve load balancing and improve service quality.

Images