Self-adaptive optimization method, device and equipment for residential electricity demand and storage medium

Abstract

The invention relates to the technical field of power systems, and provides a self-adaptive optimization method, device and equipment for residential electricity demands and a storage medium, and the method comprises the steps: obtaining multi-source data of a power distribution network, carrying out the data preprocessing and feature screening, and generating a multi-factor coupling time series data set; extracting a multi-dimensional time sequence feature vector corresponding to the multi-factor coupling time sequence data set; performing dynamic adjustment based on the multi-dimensional time sequence feature vector through a preset multi-agent asynchronous deep learning model, and generating a corresponding adaptive scheduling strategy; and performing adaptive optimization on the residential electricity demand based on an adaptive scheduling strategy. According to the invention, by providing a power demand management scheme which comprehensively considers multivariable factors and dynamically adapts to complex scenes, the operation efficiency and economical efficiency of a power system can be effectively improved, and meanwhile, the real-time power demand of residential users is met.

Description

Technical Field

[0001] This application relates to the technical field of power systems, and particularly to an adaptive optimization method, device, equipment, and storage medium for residential electricity demand. Background Art

[0002] With the rapid growth of residential electricity demand, the operating efficiency and energy utilization efficiency of power systems face new challenges. Residential electricity demand has significant seasonal, random, and hidden characteristics, which make the prediction and management of electricity demand complex and dynamic.

[0003] Currently, existing electricity demand management methods mainly rely on statistical analysis and simple time series models, such as autoregressive moving average model (ARMA) and autoregressive integrated moving average model (ARIMA). Although existing technologies can capture the linear trend of electricity demand to a certain extent, they have obvious limitations in dealing with complex non-linear relationships and multi-variable coupling characteristics. In addition, in the face of complex scenarios such as weather changes, holidays, and peak usage periods of electrical equipment, existing technologies often cannot monitor and predict the dynamic changes of residential electricity demand in a timely manner, which not only affects the optimal allocation of power resources but also may lead to the exacerbation of load peaks and energy waste. Especially during peak usage periods of electrical equipment, existing technologies are difficult to achieve dynamic load balancing and cannot meet the electricity demand of residential users at different times.

[0004] The foregoing description is for the purpose of providing general background information and does not necessarily constitute prior art. Summary of the Invention

[0005] Based on this, in order to address at least one of the above-mentioned problems, this application provides an adaptive optimization method, device, equipment, and storage medium for residential electricity demand, which can achieve the adaptive optimization of residential electricity demand, effectively improve the operating efficiency and economy of power systems, and at the same time meet the real-time electricity demand of residential users.

[0006] In a first aspect, this application provides an adaptive optimization method for residential electricity demand, including the following steps: Obtain multi-source data of the distribution network, perform data preprocessing and feature screening, and generate a multi-factor coupled time series data set; Extract the multi-dimensional time series feature vectors corresponding to the multi-factor coupled time series data set; Perform dynamic adjustment based on the multi-dimensional time series feature vectors through a preset multi-agent asynchronous deep learning model to generate corresponding adaptive scheduling strategies; Adaptively optimize the residential electricity demand based on the adaptive scheduling strategy.

[0007] Further, in some embodiments of the present application, obtaining multi-source data of the distribution network, performing data preprocessing and feature screening, and generating a multi-factor coupled time series dataset includes: Unifying the formats and cleaning the obtained multi-source data to obtain the multi-source data after data preprocessing; wherein, the multi-source data includes residential electricity load time series data, meteorological data, holiday data, and residential electricity consumption data; Using the correlation analysis method to perform feature screening on the multi-source data after data preprocessing, and generating a multi-factor coupled time series dataset based on the key factors of load demand screened out.

[0008] Further, in some embodiments of the present application, using the correlation analysis method to perform feature screening on the multi-source data after data preprocessing, and generating a multi-factor coupled time series dataset based on the key factors of load demand screened out includes: Constructing a multi-factor correlation analysis model based on a kernel function, and calculating the correlation coefficients between each load demand influencing factor and the load demand through the multi-factor correlation analysis model; Comparing the correlation coefficients with a preset correlation threshold to obtain a comparison result; Based on the comparison result, screening out the key factors of load demand; Based on the key factors of load demand, dynamically constructing a multi-factor coupled time series dataset.

[0009] Further, in some embodiments of the present application, extracting the multi-dimensional time series feature vectors corresponding to the multi-factor coupled time series dataset includes: Performing time series data enhancement processing on the multi-factor coupled time series dataset, and the time series data enhancement processing includes noise injection based on trend-cycle decomposition and random recombination of time series slices; Performing local feature extraction on the multi-factor coupled time series dataset after time series data enhancement processing through a convolutional neural network to obtain local feature vectors; Performing global feature extraction on the local feature vectors through an encoder to generate global feature vectors; Fusing the local feature vectors and the global feature vectors to generate multi-dimensional time series feature vectors including global relevance and local context.

[0010] Further, in some embodiments of the present application, dynamically adjusting based on the multi-dimensional time series feature vectors through a preset multi-agent asynchronous deep learning model to generate a corresponding adaptive scheduling strategy includes: Constructing a multi-agent asynchronous deep learning model based on the asynchronous advantage actor-critic algorithm, and initializing the global actor network and critic network; Independent agent models corresponding to different electrical devices are constructed respectively. The state space of the independent agent model includes real-time power, electricity price signals, and the probability distribution of historical electricity consumption habits. Each independent agent model calculates gradients based on the collected interaction data, updates the local network, and asynchronously shares the gradients into the global network. The global network updates the global parameters according to the shared gradients and synchronizes the updated global parameters into the local networks corresponding to each independent agent model to generate an adaptive scheduling strategy.

[0011] Furthermore, in some embodiments of the present application, the adaptive optimization of residential electricity demand based on the adaptive scheduling strategy includes at least one of the following: Generate penalties or rewards according to the difference value between the actual start-stop times and the target times of electrical devices. Generate a total cost reward according to the difference value of electricity costs before and after optimization. Dynamically adjust the penalty coefficient during peak hours and the reward coefficient during off-peak hours.

[0012] Furthermore, in some embodiments of the present application, after the adaptive optimization of residential electricity demand based on the adaptive scheduling strategy, the method further includes: Real-time monitor the actual electricity consumption data after optimization. Calculate the error value between the actual electricity consumption data and the predicted electricity consumption data corresponding to the adaptive scheduling strategy. Based on the error value, adjust the model parameters of the multi-agent asynchronous deep learning model to optimize the adaptive scheduling strategy.

[0013] In a second aspect, the present application provides an apparatus for adaptive optimization of residential electricity demand, including: A data acquisition module, configured to acquire multi-source data of a distribution network, perform data preprocessing and feature screening, and generate a multi-factor coupled time series dataset. A feature extraction module, configured to extract multi-dimensional time series feature vectors corresponding to the multi-factor coupled time series dataset. A strategy generation module, configured to perform dynamic adjustment based on the multi-dimensional time series feature vectors through a preset multi-agent asynchronous deep learning model, and generate a corresponding adaptive scheduling strategy. A demand optimization module, configured to perform adaptive optimization of residential electricity demand based on the adaptive scheduling strategy.

[0014] In a third aspect, the present application further provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the adaptive optimization method for residential electricity demand as described in the first aspect are implemented.

[0015] In a fourth aspect, the present application further provides a computer-readable storage medium storing a computer program, which when executed by a processor implements the steps of the adaptive optimization method for residential electricity demand as described in the first aspect.

[0016] As described above, the present application provides an adaptive optimization method, device, equipment, and storage medium for residential electricity demand. First, by obtaining multi-source data of the distribution network and performing data preprocessing and feature screening, a multi-factor coupled time series dataset is generated, solving the limitations of traditional methods that only rely on a single data source or a simple time series model, and being able to comprehensively consider multiple coupled factors to more comprehensively capture the dynamic changes of electricity demand. Secondly, by extracting multi-dimensional time series feature vectors, not only can the complex non-linear relationships in time series data be captured, but also the global relevance and local context information are retained, improving the prediction accuracy of the model for electricity demand. Finally, based on the dynamic adjustment strategy of the multi-agent asynchronous deep learning model, the changes in residential electricity consumption behavior and external characteristics are monitored in real time, and the scheduling strategy is optimized through the reinforcement learning algorithm. This can not only smooth the load fluctuation, reduce the electricity cost, but also dynamically adjust the strategy according to real-time data to meet the electricity demand of residential users at different times. It can be seen that the present application can effectively solve the limitations of traditional electricity demand management methods in complex scenarios, achieve the adaptive optimization of residential electricity demand, improve the operation efficiency and energy utilization efficiency of the power system, while reducing the load peak and avoiding energy waste. Description of the Drawings

[0017] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0018] Figure 1 It is an application environment diagram of the adaptive optimization method for residential electricity demand provided by the embodiment of the present application; Figure 2 It is a schematic flowchart of the adaptive optimization method for residential electricity demand provided by the embodiment of the present application; Figure 3 It is a schematic structural diagram of the adaptive optimization device for residential electricity demand provided by the embodiment of the present application; Figure 4 It is a schematic structural diagram of an adaptive optimization system for residential electricity demand provided by an embodiment of the present application; Figure 5 It is a schematic structural diagram of a computer device provided by an embodiment of the present application. Detailed implementation manners

[0019] Here, exemplary embodiments will be described in detail, and examples thereof are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of systems and methods consistent with the examples detailed in the appended claims or some aspects of the present application.

[0020] It should be noted that in this document, descriptions such as "including", "comprising" or any other variant thereof are intended to cover non-exclusive inclusion. Thus, a process, method, article or device including a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, method, article or device including that element. In addition, components, features, and elements with the same name in different embodiments of the present application may have the same meaning or different meanings, and their specific meanings need to be determined based on their explanations in the specific embodiments or further in combination with the context of the specific embodiments.

[0021] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0022] In subsequent descriptions, the use of suffixes such as "module", "component" or "unit" to represent elements is only for the convenience of describing the present application, and it has no specific meaning itself. Therefore, "module", "component" or "unit" can be used interchangeably.

[0023] The residential electricity demand has the characteristics of "seasonality, fluidity, and concealment". Currently, the management of the distribution transformer area relies on experience and the large-scale employment of personnel, making it difficult to monitor and predict the complex changes in residential electricity consumption in a timely manner. The accuracy of supporting the handling of electricity problems is insufficient, and the foresight is lacking, making it impossible to support the discovery of the laws of residential electricity demand. Facing multiple coupled and superimposed characteristics and being unable to reasonably predict the complex and changeable electricity demand, traditional electricity demand management methods often rely on statistical methods or simple time series analysis, such as constructing autoregressive moving average models and autoregressive integrated moving average models for analysis. These practices of constructing regression models based on time series can only capture linear trends, are difficult to handle non-linear relationships, and cannot directly introduce exogenous variables, affecting the explanatory power of the models and making it difficult to cope with the dynamic changes in electricity demand. Especially in situations such as weather changes, holidays, and peak usage periods of electrical equipment, traditional methods often cannot fully optimize the utilization efficiency of electricity resources, resulting in energy waste.

[0024] To solve the above technical problems, the present application provides an adaptive optimization method, device, equipment, and storage medium for residential electricity demand, which can achieve the adaptive optimization of residential electricity demand, effectively improve the operating efficiency and economy of the power system, and at the same time meet the real-time electricity demand of residential users.

[0025] Figure 1 It is an application environment diagram of the adaptive optimization method for residential electricity demand in an embodiment. Refer to Figure 1 , the adaptive optimization method for residential electricity demand is applied to the adaptive optimization system for residential electricity demand. The adaptive optimization system for residential electricity demand includes a terminal 110 and a server 120. The terminal 110 and the server 120 are connected through a network. The terminal 110 can specifically be a desktop terminal or a mobile terminal, and the mobile terminal can specifically be at least one of a mobile phone, a tablet computer, a laptop computer, etc. The server 120 can be implemented by an independent server or a server cluster composed of multiple servers. The terminal 110 is used to obtain multi-source data of the distribution network, perform data preprocessing and feature screening to generate a multi-factor coupled time series data set; extract multi-dimensional time series feature vectors corresponding to the multi-factor coupled time series data set; perform dynamic adjustment based on the multi-dimensional time series feature vectors through a preset multi-agent asynchronous deep learning model to generate corresponding adaptive scheduling strategies; and perform adaptive optimization on the residential electricity demand based on the adaptive scheduling strategies.

[0026] Please refer to Figure 2 , Figure 2 is a schematic flow diagram of the adaptive optimization method for residential electricity demand provided by an embodiment of the present application. The adaptive optimization method for residential electricity demand can specifically include the following steps: S1. Obtain multi-source data of the distribution network, perform data preprocessing and feature screening to generate a multi-factor coupled time series data set; Specifically, for step S1, by obtaining multi-source data of the distribution network (such as residential electricity load time series data, meteorological data, holiday data, and residential electricity equipment characteristic data), and performing preprocessing and feature screening on it, a time series dataset containing multi-factor coupling is finally generated. The acquisition and preprocessing of multi-source data are the basis of the entire method, ensuring that subsequent feature extraction and optimization strategies can be based on high-quality data.

[0027] In a specific embodiment, for the sources of multi-source data, it can specifically include 96-point high-frequency load data, historical meteorological data (temperature, precipitation, humidity, etc.), holiday data (such as Spring Festival, National Day, etc.), and residential electricity equipment characteristic data (such as the operating status of high-power electrical appliances such as air conditioners, charging facilities, and water heaters). For data preprocessing, the original data can be specifically unified in format, data cleaning (removing noise and outliers), and normalization processing to ensure the consistency and availability of the data. Feature screening is to screen out the key factors that have the greatest impact on load demand (such as temperature, holiday type, equipment operating status, etc.) through correlation analysis or kernel function models, and construct a time series dataset with multi-factor coupling.

[0028] This step comprehensively captures the dynamic changes of residential electricity demand by integrating multiple data sources, avoiding prediction biases caused by single data sources; by data preprocessing and feature screening, redundant and noisy data are removed, improving the training efficiency and prediction accuracy of subsequent models; through the real-time acquisition and processing of multi-source data, real-time data support is provided for subsequent dynamic optimization.

[0029] S2. Extract multi-dimensional time series feature vectors corresponding to the multi-factor coupling time series dataset; Specifically, for step S2, through feature extraction techniques, multi-dimensional feature vectors are extracted from the multi-factor coupling time series dataset, including local features (such as short-term fluctuations) and global features (such as long-term trends), and a comprehensive feature vector is generated through fusion. For example, a convolutional neural network (CNN) is used to extract local features from the time series data to capture short-term fluctuations and periodic features; an encoder (such as a Transformer encoder) is used to extract global features from the local features to capture long-term trends and context information; finally, the local features and global features are fused to generate a multi-dimensional feature vector containing global relevance and local context information. The multi-dimensional feature vector provided in this step can capture complex non-linear relationships in time series data, improving the model's prediction ability for electricity demand; through feature fusion, the model is more robust to noise and outliers and can more stably handle complex scenarios; feature extraction techniques can enhance the model's generalization ability, making it perform better in different environments and scenarios.

[0030] S3. Dynamically adjust based on the multi-dimensional time series feature vector through a preset multi-agent asynchronous deep learning model to generate a corresponding adaptive scheduling strategy; Specifically, for step S3, this step uses a multi-agent asynchronous deep learning model (such as the A3C algorithm) to dynamically adjust the scheduling strategy according to the extracted multi-dimensional time series feature vector. Among them, each agent runs independently and shares gradients, and optimizes the scheduling strategy through reinforcement learning. Each agent corresponds to an electrical device (such as an air conditioner, a charging facility, a water heater), runs independently and shares gradients to achieve distributed optimization. The asynchronous advantage actor-critic algorithm (A3C) is adopted. The actor network outputs the action probability, and the critic network evaluates the action value to dynamically adjust the strategy. Finally, according to the probability distribution of real-time data and historical electricity consumption habits, the penalty coefficient during peak hours and the reward coefficient during valley hours are dynamically adjusted. This step can respond to the changes in residents' electricity consumption demand in real time through dynamically adjusting the scheduling strategy, and achieve dynamic balance of the load; through optimizing the scheduling strategy, reduce electricity costs, achieve peak shaving and valley filling, and reduce energy waste; the multi-agent architecture can adapt to the operating characteristics of different devices, and improve the adaptability and robustness of the overall scheduling strategy.

[0031] S4. Adaptively optimize the residents' electricity consumption demand based on the adaptive scheduling strategy; Specifically, for step S4, optimize the residents' electricity consumption demand by executing the adaptive scheduling strategy. The optimization objectives include reducing electricity costs, smoothing load fluctuations, and meeting the real-time electricity consumption needs of residents. For example, adjust the start and stop times of devices according to the scheduling strategy, such as starting the charging facility during valley hours and reducing the running time of the air conditioner during peak hours. Through reward mechanisms (such as electricity cost difference rewards, start and stop times rewards, etc.), encourage users to adjust their electricity consumption behaviors. In addition, the actual electricity consumption data after optimization can be monitored in real time, calculate the error value and feedback it to the model to further optimize the scheduling strategy. This step reduces the electricity costs of residents through optimizing the scheduling strategy, maximizes economic benefits; smooths load fluctuations, reduces the load pressure during peak hours, and improves the operating efficiency of the power system, so as to meet the real-time electricity consumption needs of residents, provide high-quality electricity services, and improve the user experience.

[0032] Furthermore, in some embodiments, step S1 "Obtain multi-source data of the distribution network, perform data preprocessing and feature screening, and generate a multi-factor coupled time series data set" can specifically include: S11. Unify the format and clean the obtained multi-source data to obtain the multi-source data after data preprocessing; among them, the multi-source data includes residential electricity load time series data, meteorological data, holiday data, and residential electricity consumption data; Specifically, for step S11, the multi-source data may specifically include at least one of residential electricity load time series data, meteorological data, holiday data, and residential electricity equipment characteristic data; perform data preprocessing on the obtained multi-source data, including format unification, data cleaning, and data normalization. Among them, format unification of data is to convert data from different sources into a unified format (such as timestamp alignment, unit unification, etc.); data cleaning is used to remove noise data, outliers, and missing values to ensure data quality; perform normalization processing on the data to make it suitable for subsequent feature extraction and model training. This step ensures that data from different sources can be seamlessly integrated, providing a unified data basis for subsequent processing; through cleaning and normalization, the accuracy and availability of the data are improved, and the error in model training is reduced; the preprocessed data can be directly used for feature extraction and model training, saving subsequent processing time.

[0033] In a specific embodiment, the multi-source data may include: 96-point high-frequency load data, which are the basic data for analyzing the adaptive optimization of residential load demand; historical meteorological data, mainly including temperature, precipitation, and humidity at different times in the region, which are important external factors affecting residential electricity load; holiday data, holidays such as weekends, New Year's Day, Spring Festival, Tomb-Sweeping Day, Labor Day, Dragon Boat Festival, and National Day will have a greater impact on residential load demand; large-power electrical appliances in the residential equipment system: including household air conditioners, charging facilities, and water heaters, which directly affect residential load demand.

[0034] S12. Use the correlation analysis method to perform feature screening on the multi-source data after data preprocessing, and generate a multi-factor coupled time series dataset based on the key factors of load demand screened out; Specifically, for step S12, use the correlation analysis method to screen out the key factors that have the greatest impact on load demand, and generate a multi-factor coupled time series dataset based on these factors. For example, use statistical methods (such as Pearson correlation coefficient, Spearman rank correlation coefficient, etc.) or machine learning methods (such as kernel function model) to calculate the correlation between each factor and load demand. According to the magnitude of the correlation coefficient, screen out the key factors that have the greatest impact on load demand (such as temperature, holiday type, equipment operation status, etc.). Dynamically combine the screened key factors to generate a multi-factor coupled time series dataset. This step screens out key factors through correlation analysis, reduces the data dimension, and improves the model training efficiency; by clarifying which factors have the greatest impact on load demand, it provides more meaningful input for subsequent model training, making the generated multi-factor coupled time series dataset more in line with actual needs and better supporting subsequent feature extraction and optimization strategies.

[0035] Further, in some embodiments, step S12, "Performing feature screening on the multi-source data after data preprocessing by using the correlation analysis method, and generating a multi-factor coupled time series data set based on the key factors of the load demand selected", may specifically include: S121. Constructing a multi-factor correlation analysis model based on a kernel function, and calculating the correlation coefficients between each load demand influencing factor and the load demand through the multi-factor correlation analysis model; Specifically, this step constructs a multi-factor correlation analysis model through a kernel function for calculating the correlation coefficients between each load demand influencing factor (such as meteorological data, holiday data, equipment operation status, etc.) and the load demand. For example, the Gaussian kernel function is used to calculate the correlation coefficient between meteorological data and the electricity load demand. The kernel function can map the data into a high-dimensional space to capture complex non-linear relationships. Additionally, linear kernels and polynomial kernels can also be selected, and the appropriate kernel function is chosen according to the data characteristics and application scenarios. The correlation coefficients between each factor and the load demand are calculated through the kernel function to quantify the influence degree of each factor on the load demand. The historical data is used to train the correlation analysis model to ensure that the model can accurately reflect the relationship between each factor and the load demand. The kernel function provided in this step can effectively capture the complex non-linear relationships between each factor and the load demand, improving the accuracy of the correlation analysis; the influence of each factor on the load demand is quantified through the correlation coefficients, providing a basis for subsequent feature screening; the correlation analysis model can adapt to different data characteristics and application scenarios and has strong generalization ability.

[0036] S122. Comparing the correlation coefficients with a preset correlation threshold to obtain a comparison result; Specifically, this step compares the calculated correlation coefficients with the preset correlation threshold to determine which factors have a significant influence on the load demand. The correlation threshold is set according to actual requirements and experience, such as 0.5 or 0.7, for screening out the factors with a greater influence on the load demand. The correlation threshold can also be dynamically adjusted according to the performance of the model and the changes in the data to ensure that the screened factors always have a significant influence. This step screens out the key factors with a greater influence on the load demand by comparing the correlation coefficients and the threshold, reducing the data dimension; reducing the amount of data for subsequent processing and improving the efficiency of model training and optimization; ensuring that the screened factors have a significant influence on the load demand and improving the prediction accuracy of the model.

[0037] S123. Based on the comparison result, screening out the key factors of the load demand; Specifically, based on the comparison results, this step filters out the key factors that have the greatest impact on the load demand, and these factors will be used for the construction of the subsequent multi-factor coupled time series dataset. According to the magnitude and threshold of the correlation coefficient, factors such as temperature, holiday type, and equipment operation status that have the greatest impact on the load demand are screened out. According to the real-time data and model performance, the screening results of the key factors are dynamically updated to ensure that the model is always optimized based on the latest key factors. This step provides more meaningful inputs for the subsequent model training by clarifying which factors have the greatest impact on the load demand; the generated multi-factor coupled time series dataset better meets the actual needs and can better support the subsequent feature extraction and optimization strategies; by screening out the key factors, the training efficiency and prediction accuracy of the model are effectively improved.

[0038] S124. Dynamically construct a multi-factor coupled time series dataset based on the key factors of load demand; Specifically, based on the screened key factors, this step dynamically constructs a multi-factor coupled time series dataset for subsequent feature extraction and optimization strategies. The screened key factors are dynamically combined to generate a multi-factor coupled time series dataset. According to the real-time data and changes in the key factors, the dataset is dynamically updated to ensure that the dataset always reflects the latest changes in the load demand. Through data fusion technology, the key factor data from different sources is integrated into a unified dataset. The multi-factor coupled time series dataset generated in this step better meets the actual needs and can better support the subsequent feature extraction and optimization strategies; the dynamically constructed dataset can reflect the changes in the load demand in real time and provide real-time data support for the subsequent dynamic optimization; by dynamically updating the dataset, it is ensured that the model is always trained and optimized based on the latest data, improving the prediction accuracy and adaptability of the model.

[0039] In a specific embodiment, since the above data is distributed in different systems and there are problems such as inconsistent field types for various types of data, resulting in discrete data distribution, it is necessary to preprocess the multi-source data, and then screen out relatively appropriate relevant factor data and conduct a preliminary evaluation of the data.

[0040] First, construct a data evaluation grading difference control function: Among them, W is the objective function, representing the quantity to be optimized, is the weight coefficient of the i-th type of influencing factor, are the i-th and j-th data points in the dataset, is the action unit of the internal structure of the data, K is the kernel function, a is the Lagrangian operator, which can dynamically predict the three elements of the multi-source data evaluation; the min in the formula indicates that the objective function W is to be minimized.

[0041] Secondly, by selecting various influencing factors that affect data evaluation, the first-order partial derivative and mean square error of the sample data W are obtained, and the quantitative condition discriminant formula for various influencing factors is calculated as follows: Wherein, is the quantitative condition discriminant formula for the i-th type of influencing factor, is the weight coefficient of the i-th type of influencing factor, reflecting the importance of this factor in the overall evaluation, K is the kernel function, and respectively represent the output values of the i-th and j-th types of influencing factors, and b is the bias term used to adjust the output of the model.

[0042] Thirdly, a multi-layer step-by-step dynamic prediction of the evaluation coefficient based on the data evaluation model is carried out to obtain the evaluation discriminant formula for multi-source data: Wherein, C represents a specific threshold for distinguishing different evaluation states, represents the state where the evaluation result is "normal", corresponding to the case where the weight coefficient is equal to 0. represents the state where the evaluation result is "risk", corresponding to the case where the weight coefficient is between 0 and C. represents the state where the evaluation result is "abnormal", corresponding to the case where the evaluation coefficient is equal to C.

[0043] Based on the above analysis, a multi-source data evaluation rule is constructed, and a substitution data method is used to fit the state characteristics of the parameters. The fitting result is as follows: Wherein, is the data result of the selected relevant factors; x(t) represents the state characteristic vector of multi-source data at time t, x0(t) represents the data result of the 0th relevant factor at time t, and so on, represents the data result of the (k - 1)-th relevant factor at time t; T represents the transpose operation of the vector, converting the row vector into a column vector.

[0044] Based on the above characteristic fitting, a canonical analysis and empirical analysis sample set for data evaluation is obtained.

[0045] Furthermore, in some embodiments, step S2, "extracting the multi-dimensional time-series feature vector corresponding to the multi-factor coupled time-series data set", may specifically include: S21. Perform time series data augmentation on the multi-factor coupled time series dataset. The time series data augmentation includes noise injection based on trend-cycle decomposition and random recombination of time series slices; Specifically, this step processes the multi-factor coupled time series dataset through time series data augmentation techniques to improve the generalization ability and robustness of the model. Among them, time series data augmentation includes noise injection based on trend-cycle decomposition and random recombination of time series slices. Decompose the time series data into trend terms, periodic terms, and residual terms, and generate new data sequences by adjusting the trend terms and periodic terms. Add noise that conforms to the normal distribution to the decomposed data to simulate the random fluctuations in the actual data. Cut the time series data into multiple subsequences and randomly recombine these subsequences to generate new data samples. This step generates more diverse data samples through data augmentation, improving the model's adaptability to different scenarios; noise injection can improve the model's robustness to data noise and avoid overfitting; the recombination of time series slices increases the diversity of the data, enabling the model to learn richer features.

[0046] S22. Extract local features from the multi-factor coupled time series dataset after time series data augmentation through a convolutional neural network to obtain local feature vectors; Specifically, this step uses a convolutional neural network (CNN) to extract local features from the augmented time series data, capturing short-term fluctuations and local patterns. Use a one-dimensional convolutional layer to extract local features from the time series data. Increase the receptive field through dilated convolution to capture more extensive context information. Retain the original input information through residual connections to accelerate model convergence. The CNN in this step can effectively extract local features from time series data, improving the model's sensitivity to short-term fluctuations; convolutional operations have high computational efficiency and are suitable for processing large-scale time series data; dilated convolution and residual connections can capture richer local features and improve the model's expressive ability.

[0047] S23. Extract global features from the local feature vectors through an encoder to generate global feature vectors; Specifically, this step extracts global features from the local feature vectors through an encoder (such as a Transformer encoder), capturing long-term trends and global context information. It includes: using a multi-head attention mechanism and a feed-forward neural network to extract global features, and integrating time series information into the feature vectors through positional encoding; masking the features at some time points to enhance the model's prediction ability. The encoder provided in this step can effectively extract global features from time series data, improving the model's sensitivity to long-term trends; the multi-head attention mechanism can capture the correlations between different time points and enhance the model's context understanding ability; the masking design improves the model's prediction ability, enabling it to better handle missing data.

[0048] S24. Fuse the local feature vector and the global feature vector to generate a multi-dimensional time-series feature vector that includes global relevance and local context; Specifically, this step fuses the local feature vector and the global feature vector to generate a multi-dimensional time-series feature vector that contains global relevance and local context information. Simple concatenation, weighted summation, or more complex fusion networks (such as attention fusion networks) can be used, and dimensionality reduction techniques (such as PCA) are used to reduce the dimensionality of the fused features and improve computational efficiency. Finally, the fused feature vector is further optimized through contrastive learning or self-supervised learning. The fused feature vector in this step contains both local and global features, which can describe time-series data more comprehensively; the multi-dimensional feature vector can improve the prediction accuracy and generalization ability of the model; through feature dimensionality reduction and optimization, the computational complexity is reduced and the running efficiency of the model is improved.

[0049] Specifically, for the fusion of the global features extracted by the Transformer encoder and the local features extracted by the CNN, the process is as follows: First, obtain the preprocessed multi-factor coupled time-series data set, which contains multi-source data such as residential electricity load time-series data, meteorological data, holiday data, and residential electricity equipment feature data. Use the one-dimensional convolutional layer of the convolutional neural network (CNN) to perform a sliding window operation on the time-series data to automatically learn local features. For example, a convolutional kernel of size 3 and a stride of 1 are used to perform a convolutional operation on the time-series data to extract short-term fluctuations and periodic features. After the convolutional operation, the ReLU activation function is used to perform a non-linear transformation on the features to enhance the expression ability of the model. The max pooling layer is used to reduce the dimensionality of the features, retain the most important feature information, and at the same time reduce the computational amount and prevent overfitting. After the above operations, a local feature vector is obtained, with a shape of (sequence length, local feature dimension).

[0050] Then, the local feature vector is combined with the positional encoding through the Transformer encoder. The positional encoding uses the sine-cosine encoding method to provide time-series information for the model. Through the multi-head self-attention mechanism of the Transformer encoder, the model captures dependencies between different positions and focuses on important feature positions to extract global features. For example, 8 attention heads are set, and the dimension of each head is 64. Non-linear transformations are performed on the features at each position to further extract high-level features. After the multi-head self-attention mechanism and the feed-forward neural network, residual connections and layer normalization are performed respectively to stabilize model training and accelerate convergence. After passing through multiple layers of the Transformer encoder, a global feature vector is obtained, with a shape of (sequence length, global feature dimension).

[0051] Next, the local feature vector and the global feature vector are concatenated in the feature dimension to obtain a fused feature vector with a shape of (sequence length, local feature dimension + global feature dimension). For example, if the local feature dimension is 128 and the global feature dimension is 256, then the fused feature dimension is 384.

[0052] Finally, a fully connected layer is used to perform a linear transformation on the fused feature vector, mapping the features to a task-related space, such as a feature space with a dimension of 512. The ReLU activation function is used to perform a non-linear transformation on the fused features to enhance the model's expressive power. A Dropout layer is added to randomly discard some neurons to prevent overfitting and improve the model's generalization ability. Finally, a feature vector that fuses local and global features is obtained as the input for the subsequent generation of the adaptive scheduling strategy. The fused feature vector is input into the adaptive scheduling strategy generation module based on the multi-agent asynchronous deep learning model to generate an optimized residential electricity demand scheduling strategy. The multi-agent asynchronous deep learning model dynamically adjusts the strategies of each agent according to the rich information in the fused feature vector to achieve the adaptive optimization of residential electricity demand.

[0053] In a specific embodiment, based on the above multi-source data acquisition and preprocessing, a time series dataset with multi-factor coupling based on 96-point high-frequency load data is obtained, and the residential load characteristics under multi-factor coupling are further extracted. First, a data augmentation strategy is performed on the time series data with multi-factor coupling, and the data is transformed. Secondly, the two newly augmented and transformed time series data are incorporated into the encoder and decoder to complete the feature transformation. Thirdly, position encoding is fused based on the Transformer encoder mask method to extract the time feature vector and perform cross-prediction on the masked part. Finally, the time feature vector after data augmentation is selected as the positive sample, and the time feature vectors of other sequences are used as negative samples. The positive and negative samples are used for learning and training. After the training is completed, the above embedded features are the results of the extraction of the original sequence features.

[0054] (1) Time series data augmentation Time series data enhancement is a method to improve the generalization ability of the model by generating new data samples, especially when the amount of data is insufficient. For example, for time series data in years, it contains 96 points of high-frequency load data, historical meteorological data, holiday data, and high-power electrical appliance data of residential equipment systems. Since the significant features of holiday data, meteorological fluctuations, and high-power electrical appliance data of residential equipment systems are relatively small compared to the total amount of data, the time series data needs to be enhanced. Traditional time series data usually have trends, periodicity, and noise. Enhanced data can help improve the generalization ability of the model and avoid overfitting. Common data enhancement strategies include data conversion, interpolation, adding artificial interference, sliding window techniques, and data synthesis. They do not consider removing the noise that may exist in the original data, but only add and delete operations on the original sequence data. The patent of this invention uses two enhancement methods. One is to decompose the time series into trend factors, periodic factors, and residual terms. By combining the trend terms and periodic terms, a sequence homologous to the original sequence is obtained, and then the residual terms are eliminated; the second is to generate new time series data by slicing and combing the time series data.

[0055] 1) PTL sequence decomposition This data enhancement method can retain the relevant features in the time series and generate different enhanced time series by adding the product of data scale that conforms to the normal distribution and random noise to the newly generated time series. The specific steps are as follows: For each sample in the time series dataset First, use the preset window size to perform PTL decomposition to obtain trend factors and cyclical factors , and then according to the preset parameters Adjust the size of the normal distribution noise as follows: Among them, is the standard deviation of the residual, i.e., the underlying size of the normally distributed noise; is the maximum value of the sample; is the minimum value of the sample; k represents the parameter used to adjust the noise scale of the normal distribution, and its value is equal to the absolute value of the difference between the maximum and minimum values ​​of the sample; It means the mean is 0 and the variance is Normally distributed noise; To adjust the noise scale after normal distribution; abs means taking the absolute value operation. By using PTL decomposition to obtain a new data sequence, the trend change is relatively significant compared to the original sequence. At the same time, by adjusting the parameters, the fluctuation of the enhanced sequence can be controlled, further improving the data enhancement effect. The fluctuation of the enhanced sequence can be controlled by adjusting the parameters, which improves the effect of data enhancement.

[0056] 2) Time series data slicing augmentation The SAN attempts to eliminate the non-stationarity of the time series by using local time series slices (i.e., subsequences) as units instead of the overall time series slice. By slicing the time series and then swapping the order of each slice to generate a new data sequence. The time series data is randomly sliced into slice data of the same length with a quantity less than P through preset parameters, and then the slice data is re-spliced. The generated new sequence retains the information of the original sequence, enabling the model to judge the specific characteristics of each slice. Finally, normal distribution noise decomposed by PTL is added to this sequence, and the specific method is as follows: where x is the trend factor and periodic factor of the time series; split represents the segmentation of the time series data; Shift represents the re-splicing of the slice data; is a hyperparameter that controls the standard deviation size of the normal distribution; P represents the maximum number of slices of the time series, and P is dynamically adjusted according to the length of the time series. When the time series length is very long, the length of the data slice can be appropriately increased to enhance the effect of the slice, and the effect of the slice can be enhanced by increasing the number of slices. If the time series length is short, the number of data splits can be reduced to reduce the deviation of the original sequence.

[0057] The new time series formed by the above two data augmentations becomes the input data of the entire feature extraction model, where k, can preset to control the noise distribution.

[0058] (2) Encoder-decoder design 1) Encoder design The encoder is composed of N residual modules stacked together. Each residual module consists of a one-dimensional convolutional layer, a dilated convolutional layer, a normalization layer, and a non-linear activation function, and the input is directly added to the output through a residual connection. The first module processes the original input, and the subsequent modules process the output of the previous module in turn, finally generating the embedded features. Through the residual connection, the input of each residual module first passes through the one-dimensional convolutional layer to extract local features; then through the dilated convolutional layer to increase the receptive field and capture a wider context. After each convolution, there is a normalization layer to stabilize the activation and accelerate training; then the results of the two-layer convolution processing are integrated, and the feature length is reduced through the max-pooling layer and the non-linear activation function; finally, it is input into the next residual module for continuous iteration. Since the convolutional layer will increase or decrease the number of channels to a certain extent, the convolutional kernel of the residual connection one-dimensional convolution in this invention patent is 1, and the one-dimensional convolution projects the input features of the time series into the number of output channels. In addition, the dilation coefficient of the dilated convolutional layer varies with the depth of the model and is proportional to the depth of the model layer, that is, the deeper the model layer, the larger the dilation coefficient of the dilated convolutional layer. After the above iteration, the output result of the last residual module is the representation of the original sequence in the embedding layer. The specific method is as follows: Where represents the output of the i-th residual module, represents the normalization and non-linear activation function operations in the residual module, and represent the weight matrix and bias term of the dilated convolution and the one-dimensional convolution, and represent the projection weight matrix and bias term of the residual connection, and i takes positive integers that increase in odd numbers.

[0059] 2) Decoder design The embedded features output by the above encoder are input into the decoder module. The decoder consists of a dilated convolutional layer and two one-dimensional convolutional layers. Among them, the dilation coefficient of the dilated convolutional layer in the decoder is symmetric with that of the dilated convolutional layer in the encoder, and the last one-dimensional convolutional layer is symmetric with the max-pooling operation in the encoder. The upsampling of the data features is achieved by adjusting the stride parameter of the one-dimensional convolution. The reconstructed sequence output by the last residual module has the same shape as the original sequence. The mean square error between the output reconstructed sequence and the original sequence without enhancement is used as the reconstruction loss to participate in the backpropagation of the reverse gradient of the model training, thereby updating the model parameters. The specific method is as follows: Where represents the i-th decoder module, represents the normalization and non-linear activation function operations, and x is the original sequence, represents the reconstructed sequence of the decoder output, and denote the weight matrices and bias terms of dilated convolution and one-dimensional convolution, MSE is the mean squared error loss function, is the reconstruction loss. Reconstructing based on the embedded representation after data augmentation and the original sequence without augmentation can improve the denoising ability of the model.

[0060] (3) Feature extraction To enable the model to better capture the features of the time series, based on the data augmentation technology described above, the present invention generates two different input sequence transformations, namely and . Then the two transformed sequences are input into the encoder to obtain the embedded features and , and after adding positional encoding, they are input into a Transformer encoder with N layers for feature extraction. Among them, the structure of each Transformer encoder consists of multi-head attention, feed-forward neural network, normalization, and residual connection. The positional encoding adopts the same sine-cosine encoding method as the Transformer model. Specifically as follows: where f is the feed-forward neural network including normalization operations, MHA is the multi-head attention mechanism, is the combination of positional encoding and embedded features, is the sine positional encoding, is the cosine positional encoding, pos is the position index in the input sequence, i is the dimension index, dmodel is the encoded dimension, sin is the sine function, and cos is the cosine function. By adding a mask design to the embedded features and , the feature values after the time point to be predicted are masked. By inputting a partial sequence of of the embedded features into the Transformer for learning, the last vector of the output of the last layer of the Transformer encoder is used as the time feature vector , which fuses the context information of the time point . Predict the from the time feature vector obtained by PTL enhancement to the T time period for the features obtained by shift enhancement, while the time feature vector obtained by shift enhancement predicts the The time period from 0 to T. The time series embedding features obtained by the above two enhancement techniques are used as positive samples, while other samples in the batch are used as negative samples for measurement. The dot product operation is performed between the predicted value and the corresponding true feature value and minimized, and the dot product operation is performed between the predicted value and the negative samples and maximized. Therefore, the contrastive learning loss can be obtained: where, is the PTL contrastive learning loss, is the shift contrastive learning loss, is the projection weight matrix at the predicted time point t, and are respectively and the time feature vectors output by the N-layer Transformer model, and are the embedding features.

[0061] Meanwhile, the time feature vectors and obtained by the above two data enhancement techniques are input into the projection module composed of a linear layer and a non-linear activation function, and the mean square error of the outputs of these two vectors is calculated as the time feature projection loss: where, is the time feature projection loss, represents the operations of the linear layer and the non-linear activation function.

[0062] (4) Definition of the loss function The loss function required for model training consists of three parts: where , , represent the weights of the three sub-loss functions in the loss function, represents the reconstruction loss, represents the contrastive learning loss based on the PTL-enhanced data, represents the contrastive learning loss based on the time series slice-enhanced data. Through continuous iterative training of the model, it learns more representative and discriminative embedding features of the time series data.

[0063] Furthermore, in some embodiments, step S3 "dynamically adjust based on the multi-dimensional time series feature vectors through a preset multi-agent asynchronous deep learning model to generate corresponding adaptive scheduling strategies" may specifically include: S31. Construct a multi-agent asynchronous deep learning model based on the asynchronous advantage actor-critic algorithm and initialize the global actor network and critic network; Specifically, this step constructs a multi-agent asynchronous deep learning model through the asynchronous advantage actor-critic algorithm (A3C) to dynamically adjust the scheduling strategy. The actor network is responsible for outputting action probabilities, and the critic network is responsible for evaluating the value of actions. The A3C algorithm is an asynchronous reinforcement learning algorithm that explores the environment in parallel through multiple agents and updates the global network parameters asynchronously to improve the learning efficiency. The actor network is used to output the action probabilities of the device at different time points to guide the start-stop decision of the device. The critic network is used to evaluate the action selection of the actor network, calculate the advantage function, and guide the update of the actor network. This step accelerates the training process of the model through the parallel operation of multiple agents; the asynchronous update mechanism avoids the synchronous waiting for model updates and improves the training efficiency; the action value is evaluated through the advantage function to optimize the policy output of the actor network.

[0064] S32. Construct corresponding independent agent models for different electrical devices respectively. The state space of the independent agent model includes real-time power, electricity price signal, and historical electricity consumption habit probability distribution; Specifically, this step constructs an independent agent model for each electrical device (such as air conditioners, charging facilities, water heaters). The independent agent model refers to an independent decision-making unit constructed for each electrical device (such as air conditioners, charging facilities, water heaters, etc.) in a multi-agent system. Each agent is responsible for making the optimal start-stop decision based on its own state and environmental information to achieve the adaptive optimization of residential electricity demand.

[0065] The composition of the independent agent model is as follows: State space: Describes the state of the environment where the agent is located, including real-time power, electricity price signal, and historical electricity consumption habit probability distribution, etc. These state information provide a basis for the agent's decision-making.

[0066] Action space: The actions that the agent can take, such as the start-stop operation of the device. The selection of actions will affect the operating state and electricity load of the device.

[0067] Policy network (Actor Network): A neural network that outputs the action probability distribution according to the current state. It guides the agent to select the optimal action in a given state.

[0068] Critic Network: It evaluates the value of the actions selected by the policy network, that is, the expected future rewards of these actions. It helps the agent judge the quality of the current policy.

[0069] Reward mechanism: It defines the goals of the agent and guides the agent to learn the optimal policy through reward signals. The rewards can be based on the reduction of electricity costs, the smoothness of load fluctuations, etc.

[0070] Gradient update: The agent calculates the gradient based on the collected interaction data for updating the local network parameters. Gradient update is a key step in optimizing the policy in reinforcement learning.

[0071] Each device corresponds to an independent agent, which operates and updates the policy independently. The state space of the independent agent includes real-time power, electricity price signals, and the probability distribution of historical electricity consumption habits, comprehensively describing the operating state and environmental characteristics of the device. Customized modeling is carried out according to the operating characteristics of the device and user habits to improve the pertinence of the policy. The independent agent provided in this step can be optimized according to the characteristics of different devices, improving the adaptability of the policy; the state space covers real-time power, electricity price signals, and historical electricity consumption habits, providing a comprehensive decision-making basis; based on the probability distribution of historical electricity consumption habits, personalized scheduling policies are generated to enhance the user experience.

[0072] S33. Each independent agent model calculates the gradient according to the collected interaction data, updates the local network, and asynchronously shares the gradient to the global network; Specifically, in this step, the gradient is calculated according to the interaction data collected by each independent agent, the local network is updated, and the gradient is asynchronously shared to the global network. The data generated by the interaction between the agent and the environment includes device status, action execution results, and reward signals. The gradient is calculated based on the interaction data to guide the update of the network parameters. The agent asynchronously shares the local gradient to the global network, and the global network integrates the gradient and updates the global parameters. This step accelerates the training process of the model by parallel computing of multiple agents; the asynchronous sharing mechanism avoids synchronous waiting and improves the training efficiency; the global network integrates the gradients of multiple agents to ensure the consistency and convergence of the global parameters.

[0073] S34. The global network updates the global parameters according to the shared gradient and synchronizes the updated global parameters to the local network corresponding to each independent agent model to generate an adaptive scheduling policy; Specifically, in this step, the global network integrates the gradients of multiple agents, updates the global parameters, and synchronizes the updated parameters to the local networks of each independent agent to generate an adaptive scheduling strategy. The global network updates the global parameters based on the shared gradients to ensure the global optimality of the parameters. The updated global parameters are synchronized to the local networks of each independent agent to ensure that all agents make decisions based on the latest parameters. Based on the updated network parameters, an adaptive scheduling strategy is generated to dynamically adjust the operating state of the devices. In this step, the global network integrates the gradients of multiple agents to ensure the global optimality of the parameters; all agents make decisions based on the latest global parameters to ensure the consistency of the strategy; the adaptive scheduling strategy can be dynamically adjusted according to real-time data to meet the real-time electricity consumption needs of residential users.

[0074] Furthermore, in some embodiments, "adaptive optimization of residential electricity demand based on the adaptive scheduling strategy" in step S4 includes at least one of the following: S41. Generate penalties or rewards based on the difference between the actual start-stop times and the target times of electrical devices. Specifically, by comparing the difference between the actual start-stop times and the target times of electrical devices, corresponding penalty or reward signals are generated. If the actual start-stop times deviate too much from the target times, a penalty signal is generated; if the actual start-stop times are close to or better than the target times, a reward signal is generated. For example, the start-stop times of each electrical device are monitored in real time, and the actual operation data is recorded; reasonable start-stop target times are set according to the device characteristics and user needs, such as the number of daily operations of an air conditioner; the difference between the actual start-stop times and the target times is calculated as the basis for generating penalties or rewards; the intensity of the penalties or rewards is dynamically adjusted to ensure the flexibility and adaptability of the strategy. The over-starting and stopping of devices are restricted through the penalty mechanism to avoid energy waste; devices are encouraged to operate according to the target times through the reward mechanism to improve energy utilization efficiency; the intensity of the penalties and rewards is dynamically adjusted to ensure that the strategy can adapt to different operating environments and user needs.

[0075] S42. Generate a total cost reward based on the difference in electricity costs before and after optimization. Specifically, by comparing the difference in electricity costs before and after optimization, a total cost reward signal is generated. If the cost after optimization is lower than the cost before optimization, a positive reward is generated; if the cost after optimization is higher than the cost before optimization, a negative reward is generated. For example, the electricity costs before and after optimization are monitored in real time, and the cost changes are recorded; the difference in costs before and after optimization is calculated as the basis for generating rewards; corresponding reward signals are generated according to the magnitude of the cost difference to encourage users to adjust their electricity consumption behavior to reduce costs; combined with historical electricity cost data, the long-term effect of the optimization strategy is evaluated. This step encourages users to reduce electricity costs through a cost reward mechanism to maximize economic benefits; dynamically adjusts the optimization strategy according to the cost difference to ensure the economy and effectiveness of the strategy.

[0076] S43. Dynamically adjust the penalty coefficient during peak hours and the reward coefficient during off-peak hours; Specifically, dynamically adjust the penalty coefficient during peak hours and the reward coefficient during off-peak hours to guide users to reduce electricity consumption during peak hours and increase electricity consumption during off-peak hours, achieving peak shaving and valley filling. For example, divide peak hours and off-peak hours according to historical data and load characteristics; dynamically adjust the penalty coefficient and reward coefficient according to the real-time load situation, such as increasing the penalty coefficient during peak hours and increasing the reward coefficient during off-peak hours. Adjust the coefficients according to real-time data to ensure that the strategy can quickly respond to load changes; guide users to adjust their electricity consumption behavior through economic incentives to achieve dynamic load balance. This step guides users to reduce electricity consumption during peak hours and increase electricity consumption during off-peak hours through dynamic coefficient adjustment to achieve dynamic load balance; reduce the load pressure during peak hours, improve the load utilization rate during off-peak hours, and enhance the operating efficiency of the power system; reduce the electricity costs of users through an economic incentive mechanism while improving the overall economy of the power system.

[0077] In a specific embodiment, the present invention is based on the Multi-Agent deep reinforcement method, combining the A3C asynchronous reinforcement learning algorithm based on action probability and value. This method dynamically optimizes the electricity consumption behavior pattern of users by real-time monitoring the changes in residents' electricity consumption behavior patterns and external characteristics, in combination with the electricity demand feature extraction technology that fuses multi-dimensional features.

[0078] (1) Asynchronous Advantage Actor-Critic algorithm By introducing an asynchronous advantage actor-critic network, this algorithm consists of a value network part and a policy network part. The value network part is where, represents the state space, which is prominently manifested as the device consumption situation, the device operating state, and the electricity price composition; refers to the value network parameters. The policy network is where, represents the policy network parameters; Denotes the action space, i.e., the working state of the device; , where denotes the state of device a at time t, denotes the average power of device a at time t, denotes the electricity price parameter at time t. When performing optimization, the Actor outputs the action probability through the round-based update system , distributing the high probability during the low-price period, thereby changing the user's electricity consumption habits. At the same time, combining the probability distribution characteristics of the user's historical electricity consumption habits analyzed above, the Actor policy output mode is parameterized and adjusted: where α + β = 1, and α and β respectively refer to the action probability selected by the system and the influence weight of the historical electricity consumption habit on the final action probability, denotes the probability distribution of the historical electricity consumption habit at time t. The Critic scores the action selection of the Actor and conducts an advantage evaluation on the feedback information of the state-action combination. Specifically as follows: where denotes the advantage function, γ ∈ [0, 1] weighs the influence degree of the future reward value on the cumulative return, denotes the reward value based on time step t + z, denotes the environmental state value function at time step t + n, denotes the evaluation of the current environmental state value function. After the Critic conducts an advantage evaluation on the feedback information, it judges and modifies the probability selected by the Actor.

[0079] The A3C algorithm coordinates the operation of multiple agents on independent CPU cores through multi-threaded parallel and asynchronous mechanisms to achieve collaborative learning of shared models and experiences. Each Agent explores independently in different environments, updates its own network parameters, and finally achieves the purpose of optimizing performance. The gradient update formula of the Agent is: where ; where φ represents the regularization parameter, denotes the policy of the entropy, and are the Agent network parameters. After the gradient update, the value network and updated policy network parameters of the Agent and . The parameter update is as follows: where and are the learning rates of the policy network and the value network respectively.

[0080] (2)Action selection for residential electricity demand According to any of the above optimization moments, multiple Agents independently make optimal decisions on air conditioners, charging facilities, and water heaters respectively to ensure that the decisions of each independent Agent reach the optimum. Then, the learned results are passed to the MainAgent to form the overall network parameters , , policy . Finally, the overall network parameters are uniformly distributed to each Agent to update the network parameters, policy parameters, and policy formulation of each Agent, so that each device reaches the overall optimal decision. The communication between the MainAgent and each Agent is realized through cooperation of multiple Agents, and each Agent can share decision-making information through the MainAgent.

[0081] The optimization goal is equivalent to maximizing the overall expected return value of the system under the optimal strategy of starting and stopping residential electricity-consuming devices: where Max represents the maximum value of E, and E is the overall expected return value of the system, refers to the total system reward value obtained in an optimization period; is the starting and stopping strategy of residential electricity-consuming devices in an optimization period.

[0082] The scheduling strategy of residential load demand is determined by the starting and stopping conditions of air conditioners, charging facilities, and water heaters, and its action execution depends on the operation probability of relevant flexible devices under specific input states. For the adaptive optimization problem solved by this invention patent, a multi-task joint reward mechanism is adopted, and this mechanism includes two reward indicators.

[0083] The first reward indicator: For all , reward vectors are set by controlling 3 types of loads. Therefore, different rewards are used for air conditioners, charging facilities, and water heaters. The following is the reward setting method for 3 Agents: where , , respectively represent the number of action executions corresponding to the flexible device, represents the daily operation target quantity of the air conditioner device, , The coefficients are obtained through microprogram experiments.

[0084] The second reward indicator: Reward setting for controlling the total residential electricity cost: Among them, c represents the electricity cost of residential users, referring to the cost after A3C optimization. The difference between c and is the potential benefit of the user.

[0085] Furthermore, in some embodiments, after adaptively optimizing the residential electricity demand based on the adaptive scheduling strategy, the method may specifically further include: S51. Real-time monitor the actual electricity consumption data after optimization; Specifically, through the real-time monitoring system, obtain the actual electricity consumption data after optimization, including information such as the operating status, power consumption, start-stop time of the equipment, etc. For data collection, collect the operating data of the equipment in real time through sensors or API interfaces, use a high-speed communication network (such as 5G, optical fiber) to ensure the real-time transmission of data, store the real-time data in an efficient database to support subsequent analysis and processing, and monitor the outliers in the real-time data to discover and handle abnormal situations in a timely manner.

[0086] S52. Calculate the error value between the actual electricity consumption data and the predicted electricity consumption data corresponding to the adaptive scheduling strategy; Specifically, this step calculates the error value between the actual electricity consumption data and the predicted electricity consumption data by comparing them, which is used to evaluate the accuracy of the optimization strategy. For example, use metrics such as mean square error (MSE), mean absolute error (MAE) to calculate the error value, dynamically evaluate the performance of the optimization strategy according to the real-time data; combine historical error data to evaluate the long-term effect of the optimization strategy, and analyze the source and influencing factors of the error to provide a basis for model optimization.

[0087] S53. Adjust the model parameters of the multi-agent asynchronous deep learning model based on the error value to optimize the adaptive scheduling strategy; Specifically, this step adjusts the parameters of the multi-agent asynchronous deep learning model according to the calculated error value to optimize the adaptive scheduling strategy. For example, adjust parameters such as the weights and learning rates of the model according to the error value; combine reinforcement learning algorithms to optimize the model parameters through reward and punishment mechanisms. In the multi-agent architecture, distributively adjust the local network parameters of each agent and asynchronously update the global network. Support online learning to adjust the model parameters in real time to adapt to the dynamically changing electricity demand.

[0088] In this embodiment, through real-time monitoring and error calculation, it is ensured that the optimization strategy is adjusted based on the latest data, improving the accuracy and real-time performance of the strategy; the model parameters are dynamically adjusted according to the error value, enabling the optimization strategy to quickly adapt to changes in electricity demand; through reinforcement learning and distributed optimization, the overall performance and stability of the model are enhanced to ensure the efficient operation of the system; the long-term effect of the strategy is evaluated by combining historical data to ensure the stability and reliability of the system.

[0089] In summary, the adaptive optimization method for residential electricity demand provided by the embodiments of this application can comprehensively consider factors such as historical load data, weather data, holiday data, and characteristics of residential electrical equipment. Combining the advantages of dilated convolution, data augmentation in contrastive learning, and autoencoders, and based on the multi-feature coupling electricity demand feature extraction technology of Transformer, it realizes feature extraction of time series information and attribute information under unlabeled data and multi-coupled features; introduces the Multi-Agent asynchronous deep reinforcement learning technology, combines the A3C asynchronous reinforcement learning algorithm based on action probability and value, and at the same time integrates it with the multi-feature coupling electricity demand feature extraction to guide residential users to reasonably arrange the usage time of electrical equipment, meet the balance and dynamic adjustment of residential users' electricity demand at different times, contribute to comprehensively improving the safe, stable, and economic operation of the distribution network, and combine with customers' electricity consumption demands to ensure high-quality electricity services.

[0090] It should be understood that although Figure 2 the steps in the flowchart are shown sequentially according to the arrows, these steps do not necessarily execute in the order indicated by the arrows. Unless clearly stated in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 2 at least a part of the steps in

[0091] can include multiple sub-steps or multiple stages. These sub-steps or stages do not necessarily execute at the same time, but can execute at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or sub-steps or stages of other steps.

[0092] Please refer to Figure 3 , Figure 3The figure is a schematic structural diagram of an adaptive optimization device for residential electricity demand provided by an embodiment of the present application. The adaptive optimization device for residential electricity demand may specifically include a data acquisition module 201, a feature extraction module 202, a policy generation module 203, and a demand optimization module 204, which are specifically as follows: The data acquisition module 201 is configured to acquire multi-source data of a distribution network, perform data preprocessing and feature screening, and generate a multi-factor coupled time series dataset; The feature extraction module 202 is configured to extract a multi-dimensional time series feature vector corresponding to the multi-factor coupled time series dataset; The policy generation module 203 is configured to perform dynamic adjustment based on the multi-dimensional time series feature vector through a preset multi-agent asynchronous deep learning model, and generate a corresponding adaptive scheduling policy; The demand optimization module 204 is configured to perform adaptive optimization on the residential electricity demand based on the adaptive scheduling policy.

[0093] For the specific limitations of the adaptive optimization device for residential electricity demand, reference may be made to the limitations of the adaptive optimization method for residential electricity demand in the above text, which will not be elaborated here. Each module in the above-mentioned adaptive optimization device for residential electricity demand can be implemented in whole or in part by software, hardware, and their combination. The above-mentioned modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to the above-mentioned modules.

[0094] The adaptive optimization device for residential electricity demand provided in this embodiment first generates a multi-factor coupled time series dataset by acquiring multi-source data of a distribution network and performing data preprocessing and feature screening, solving the limitations of only relying on a single data source or a simple time series model in traditional methods, being able to comprehensively consider various coupling factors, and more comprehensively capturing the dynamic changes of electricity demand; secondly, by extracting a multi-dimensional time series feature vector, it can not only capture the complex non-linear relationships in time series data, but also retain the global correlation and local context information, improving the prediction accuracy of the model for electricity demand; finally, based on the dynamic adjustment strategy of the multi-agent asynchronous deep learning model, it can monitor the changes in residential electricity consumption behavior and external features in real time, and optimize the scheduling strategy through a reinforcement learning algorithm, which can not only smooth the load fluctuation, reduce the electricity cost, but also dynamically adjust the strategy according to real-time data to meet the electricity demand of residential users at different times.

[0095] In a specific embodiment, as Figure 4 shown, the embodiment of the present application also provides a deep learning-based adaptive optimization system for residential load demand, including: (1) A data processing module This module includes data acquisition layer, data transmission layer, data preprocessing platform, and data analysis platform. Based on the data evaluation classification difference control function, it performs data control and data review on the above functional modules. At the same time, it obtains the multi-source data evaluation discriminant by the evaluation coefficients and the reverse multi-layer step-by-step dynamic prediction corresponding to the multi-source data evaluation model, forms the interactive coordination of multi-source data, and forms a multi-source data management evaluation and assessment system.

[0096] (2) Feature extraction module The execution of the feature extraction system mainly includes the whole process of building a feature tree, storing data, and carrying out feature extraction and arranging data. First, building a feature number can divide the extracted feature body into multiple levels to meet the needs of a certain number level and the use of mining and disposal permissions, so as to facilitate the database to smoothly receive feature data and ensure the operation of subsequent commands. Secondly, the storage of data to be mined is an important part of the feature extraction module. The mining nodes in the processing layer can directly extract parameters that meet the required feature information in the data supply environment. The data storage layer can integrate information parameters to form a predetermined feature extraction form. The feature collection layer receives the commands of the storage layer and arranges various data nodes in an orderly manner. Thirdly, feature extraction is the last step of the system execution environment. This module mainly extracts data to be stored, and the extraction standard can be formulated according to the execution requirements. In the entire extraction process, ensure that the feature number has a sufficient amount of data, and then combine the storage structure to obtain the data in the drive, and combine the above-mentioned feature extraction method to complete the extraction step. In the extraction process, it is necessary to incorporate the extraction architecture into all the information to be mined, and form a new structure of the extraction step by combing and merging. At the same time, it has strong execution power.

[0097] (3) Load demand adaptive optimization module Based on the external data or resident equipment data collected in real time through sensors or API structures, the Transformer encoder fuses the position code in a masked manner, extracts the time feature vector and cross-predicts the masked part to complete the feature extraction. Based on Multi-Agent deep reinforcement technology, multiple agents independently make optimal decisions for air conditioners, charging facilities, and water heaters to ensure that the decision of each independent agent is optimal. Then a distributed optimization framework is used for multi-objective optimization to adaptively adjust parameters such as learning rate and penalty coefficient. Based on the communication mechanism, asynchronous communication is carried out through message queues or APIs to maintain efficient collaboration. The adaptive control module dynamically adjusts the weight according to load changes, increases the penalty coefficient during peak hours, and reduces the penalty coefficient during valley hours to achieve dynamic balance of residents' electricity consumption.

[0098] In this embodiment, a load demand feature extraction technology based on multi-factor coupling of Transformer is used to extract features from unlabeled data, time-series information, and attribute information under multiple coupling factors. The Multi-Agent asynchronous deep reinforcement learning technology is introduced, which combines the A3C asynchronous reinforcement learning algorithms based on action probability and value. It is integrated with the load demand feature extraction of multi-factor coupling, combines external factors and the actual electricity consumption demand characteristics of residents, guides residents to reasonably arrange the start and stop of equipment, and can meet the real-time scheduling needs of residents. Through the analysis of the probability distribution characteristics of historical electricity consumption habits of residents in different environments, the adaptive optimization adjustment of residents' load demand is realized.

[0099] In addition, an embodiment of the present application also provides an electronic device, as Figure 5 shown, which shows a schematic structural diagram of the electronic device involved in the embodiment of the present application. Specifically: The electronic device may include a processor 301 with one or more processing cores, a memory 302 with one or more computer-readable storage media, a power supply 303, an input unit 304, and other components. Those skilled in the art can understand that Figure 5 the structure of the electronic device shown in does not constitute a limitation on the electronic device, and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Among them: The processor 301 is the control center of the electronic device, connects various parts of the entire electronic device through various interfaces and lines, runs or executes software programs and / or modules stored in the memory 302, and calls data stored in the memory 302 to execute various functions of the electronic device and process data, thereby monitoring the electronic device as a whole. Optionally, the processor 301 may include one or more processing cores; preferably, the processor 301 may integrate an application processor and a modulation and demodulation processor. Among them, the application processor mainly processes the operating system, user interface, application programs, etc., and the modulation and demodulation processor mainly processes wireless communication. It can be understood that the above modulation and demodulation processor may not be integrated into the processor 301.

[0100] The memory 302 can be used to store software programs and modules. The processor 301 executes various functional applications and the adaptive optimization method for residential electricity demand by running the software programs and modules stored in the memory 302. The memory 302 may mainly include a program storage area and a data storage area. Among them, the program storage area can store the operating system, application programs required for at least one function (such as the sound playback function, the image playback function, etc.); the data storage area can store data created according to the use of the electronic device. In addition, the memory 302 may include high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices. Correspondingly, the memory 302 may also include a memory controller to provide the processor 301 with access to the memory 302.

[0101] The electronic device further includes a power supply 303 for supplying power to each component. Preferably, the power supply 303 can be logically connected to the processor 301 through a power management system, so as to realize functions such as management of charging, discharging, and power consumption management through the power management system. The power supply 303 may also include any components such as one or more DC or AC power supplies, a recharge system, a power failure detection circuit, a power converter or inverter, and a power status indicator.

[0102] The electronic device may further include an input unit 304, which can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function controls.

[0103] Although not shown, the electronic device may further include a display unit, etc., which will not be elaborated here. Specifically, in this embodiment, the processor 301 in the electronic device will load the executable files corresponding to the processes of one or more application programs into the memory 302 according to the following instructions, and the processor 301 will run the application programs stored in the memory 302 to implement various functions as follows: Obtain multi-source data of the distribution network, perform data preprocessing and feature screening to generate a multi-factor coupled time series dataset; extract multi-dimensional time series feature vectors corresponding to the multi-factor coupled time series dataset; perform dynamic adjustment based on the multi-dimensional time series feature vectors through a preset multi-agent asynchronous deep learning model to generate corresponding adaptive scheduling strategies; perform adaptive optimization on residential electricity demand based on the adaptive scheduling strategies.

[0104] For the specific implementation of the above operations, reference can be made to the previous embodiments, which will not be elaborated here.

[0105] In the embodiments of the present application, by obtaining multi-source data of the distribution network, performing data preprocessing and feature screening, and generating a multi-factor coupled time series dataset, the limitations of relying only on a single data source or a simple time series model in traditional methods are solved. It can comprehensively consider various coupling factors and more comprehensively capture the dynamic changes in electricity consumption demand. By extracting multi-dimensional time series feature vectors, it can not only capture the complex non-linear relationships in time series data, but also retain the global relevance and local context information, improving the prediction accuracy of the model for electricity consumption demand. Based on the dynamic adjustment strategy of the multi-agent asynchronous deep learning model, it can monitor the changes in residents' electricity consumption behavior and external features in real time, and optimize the scheduling strategy through the reinforcement learning algorithm. It can not only smooth the load fluctuations, reduce the electricity cost, but also dynamically adjust the strategy according to real-time data to meet the electricity consumption needs of residential users at different times.

[0106] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructions, or by controlling relevant hardware through instructions. The instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0107] Therefore, the embodiments of the present application provide a storage medium, which stores multiple instructions that can be loaded by a processor to execute the steps in any of the adaptive optimization methods for residential electricity consumption demand provided by the embodiments of the present application. For example, the instructions can execute the following steps: Obtain multi-source data of the distribution network, perform data preprocessing and feature screening, and generate a multi-factor coupled time series dataset; extract multi-dimensional time series feature vectors corresponding to the multi-factor coupled time series dataset; perform dynamic adjustment based on the multi-dimensional time series feature vectors through a preset multi-agent asynchronous deep learning model to generate a corresponding adaptive scheduling strategy; perform adaptive optimization on the residential electricity consumption demand based on the adaptive scheduling strategy.

[0108] For the specific implementation of each of the above operations, reference can be made to the previous embodiments, which will not be elaborated here.

[0109] Among them, the storage medium may include: read-only memory (ROM, Read Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, etc.

[0110] Since the instructions stored in the storage medium can execute the steps in any of the adaptive optimization methods for residential electricity consumption demand provided by the embodiments of the present application, the beneficial effects that can be achieved by any of the adaptive optimization methods for residential electricity consumption demand provided by the embodiments of the present application can be realized. For details, reference can be made to the previous embodiments, which will not be elaborated here.

[0111] The above has introduced in detail an adaptive optimization method, device, equipment and storage medium for residential electricity demand provided by the embodiments of the present application. Specific examples are used in this article to elaborate on the principle and implementation manner of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. An adaptive optimization method for residential electricity demand, characterized in that Including: Obtain multi-source data of the distribution network, perform data preprocessing and feature screening, and generate a multi-factor coupled time-series dataset; Extract the multi-dimensional time-series feature vectors corresponding to the multi-factor coupled time-series dataset; Based on the multi-dimensional time-series feature vectors, perform dynamic adjustment through a preset multi-agent asynchronous deep learning model to generate corresponding adaptive scheduling strategies; Based on the adaptive scheduling strategies, perform adaptive optimization on the residential electricity demand.

2. The adaptive optimization method for residential electricity demand according to claim 1, wherein The obtaining multi-source data of the distribution network, performing data preprocessing and feature screening, and generating a multi-factor coupled time-series dataset includes: Unify the formats and clean the obtained multi-source data to obtain the multi-source data after data preprocessing; wherein, the multi-source data includes residential electricity load time-series data, meteorological data, holiday data, and residential electricity consumption data; Use the correlation analysis method to perform feature screening on the multi-source data after data preprocessing, and generate a multi-factor coupled time-series dataset based on the key factors of load demand selected.

3. The adaptive optimization method for residential electricity demand according to claim 2, wherein The using the correlation analysis method to perform feature screening on the multi-source data after data preprocessing, and generating a multi-factor coupled time-series dataset based on the key factors of load demand selected includes: Construct a multi-factor correlation analysis model based on a kernel function, and calculate the correlation coefficients between each load demand influencing factor and the load demand through the multi-factor correlation analysis model; Compare the correlation coefficients with a preset correlation threshold to obtain a comparison result; Based on the comparison result, select the key factors of load demand; Based on the key factors of load demand, dynamically construct a multi-factor coupled time-series dataset.

4. The adaptive optimization method for residential electricity demand according to claim 1, wherein The extracting the multi-dimensional time-series feature vectors corresponding to the multi-factor coupled time-series dataset includes: Perform time-series data augmentation on the multi-factor coupled time-series dataset, and the time-series data augmentation includes noise injection based on trend-cycle decomposition and random recombination of time-series slices; Extract local features from the multi-factor coupled time-series dataset after time-series data augmentation through a convolutional neural network to obtain local feature vectors; Extract global features from the local feature vectors through an encoder to generate global feature vectors; Fuse the local feature vectors and the global feature vectors to generate multi-dimensional time-series feature vectors including global relevance and local context.

5. The adaptive optimization method for residential electricity demand according to claim 1, wherein The performing dynamic adjustment through a preset multi-agent asynchronous deep learning model based on the multi-dimensional time-series feature vectors to generate corresponding adaptive scheduling strategies includes: Construct a multi-agent asynchronous deep learning model based on the asynchronous advantage actor-critic algorithm, and initialize the global actor network and critic network; Construct corresponding independent agent models for different electrical appliances respectively, and the state space of the independent agent models includes real-time power, electricity price signals, and historical electricity consumption habit probability distributions; Each independent agent model calculates gradients and updates the local network according to the collected interaction data, and asynchronously shares the gradients into the global network; The global network updates the global parameters according to the shared gradients and synchronizes the updated global parameters to the local networks corresponding to each of the independent agent models to generate an adaptive scheduling strategy.

6. The adaptive optimization method for residential electricity demand according to claim 1, wherein The adaptive optimization of the residential electricity demand based on the adaptive scheduling strategy includes at least one of the following: Generating penalties or rewards according to the difference value between the actual start-stop times and the target times of the electrical equipment; Generating a total cost reward according to the difference in electricity costs before and after optimization; Dynamically adjusting the penalty coefficient during peak hours and the reward coefficient during off-peak hours.

7. The adaptive optimization method for residential electricity demand according to claim 1, characterized in that, After the adaptive optimization of the residential electricity demand based on the adaptive scheduling strategy, the method further includes: Real-time monitoring of the actual electricity consumption data after optimization; Calculating the error value between the actual electricity consumption data and the predicted electricity consumption data corresponding to the adaptive scheduling strategy; Adjusting the model parameters of the multi-agent asynchronous deep learning model based on the error value to optimize the adaptive scheduling strategy.

8. An adaptive optimization device for residential electricity demand, characterized in that, Including: A data acquisition module for acquiring multi-source data of the distribution network, performing data preprocessing and feature screening, and generating a multi-factor coupled time series dataset; A feature extraction module for extracting multi-dimensional time series feature vectors corresponding to the multi-factor coupled time series dataset; A strategy generation module for dynamically adjusting based on the multi-dimensional time series feature vectors through a preset multi-agent asynchronous deep learning model to generate a corresponding adaptive scheduling strategy; A demand optimization module for adaptively optimizing the residential electricity demand based on the adaptive scheduling strategy.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the adaptive optimization method for residential electricity demand according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the adaptive optimization method for residential electricity demand according to any one of claims 1 to 7.

Images