A neural network-based customer electricity consumption feature analysis method and system

Abstract

The application discloses a kind of based on neural network's customer electricity characteristic analysis method and system, comprising: constructing electricity chart structure, establish the directed association from unstructured event data to subsequent time series electricity data, and adopt graph neural network, time difference between event and time series data is generated as influence factor Fusion feature vector;Then, adopt hierarchical analysis neural network architecture, short time microcosmic mode in time series data is parsed into long cycle macroscopic behavior, and combine the fusion feature vector, generate deep behavior representation;Finally, based on the deep behavior representation generates dynamic customer portrait, guarantees its stability of update by state smoothing algorithm, and benefit estimation is carried out before executing business action, to realize conditional execution.The application overcomes the defects of difficult data fusion, inaccurate portrait and high decision risk in the prior art, improves the precision, stability and automatic decision level of customer feature analysis.

Description

Technical Field

[0001] This invention belongs to the field of electricity consumption analysis technology, and more specifically, relates to a method and system for analyzing customer electricity consumption characteristics based on neural networks. Background Technology

[0002] With the deepening of smart grid construction and the widespread adoption of new power-consuming facilities such as smart meters, distributed energy sources, and charging piles, the data collected by the power system is showing significant characteristics of massive growth and high-dimensional complexity. How to effectively utilize power big data to achieve accurate insights and intelligent management of user electricity consumption behavior has become a core issue in promoting the digital transformation of the power grid and improving the quality of power supply services and operational efficiency.

[0003] However, existing data analytics technologies face numerous challenges in fully utilizing this data. On the one hand, while traditional analytical models (such as clustering algorithms) can achieve preliminary user segmentation, their granularity is coarse, severely lacking the ability to finely and dynamically track the electricity consumption behavior of individual users. Consequently, they cannot accurately capture pattern fluctuations caused by seasonal changes, work schedules, or equipment changes. On the other hand, these analytical technologies are often disconnected from downstream business scenarios (such as demand response and personalized services), failing to form a data-driven automated closed loop and greatly limiting the conversion of their data value.

[0004] This inherent deficiency at the analytical level directly leads to the inherent flaws of the current customer profiling system. Existing profile tags are static, mostly limited to superficial attributes such as industry affiliation and contracted capacity, lacking the ability to dynamically depict deeper behavioral patterns such as device start / stop preferences and seasonal sensitivity. Furthermore, the long profile update cycle and low feature extraction efficiency cause business insights to generally lag behind actual user needs. The end result is that personalized services generated based on such static, coarse-grained profiles (such as energy efficiency recommendations and electricity pricing packages) have a low match with user needs, leading to poor recommendation hit rates and difficulty in effectively stimulating user engagement.

[0005] In summary, existing power big data analytics technologies have shortcomings in dynamic feature extraction, multi-source data fusion, real-time profile updates, and application linkage, making it difficult to effectively support the refined operation and personalized service needs of the power grid. Therefore, there is an urgent need to develop a novel technical solution to overcome the aforementioned deficiencies of existing technologies. Summary of the Invention

[0006] To address the challenges of multi-source data fusion, inaccurate dynamic feature extraction, unstable profile updates, and high application decision-making risks in existing technologies, this invention provides a neural network-based method and system for analyzing customer electricity consumption characteristics. The method includes: constructing an electricity consumption graph structure; establishing a directed association from unstructured event data to subsequent time-series electricity consumption data; and using a graph neural network to generate a fused feature vector by using the time difference between events and time-series data as an influencing factor. Next, a hierarchical parsing neural network architecture is used to analyze short-term micro-patterns in the time-series data into long-term macro-behaviors, and combined with the fused feature vector to generate a deep behavioral representation. Finally, a dynamic customer profile is generated based on the deep behavioral representation. A state smoothing algorithm ensures the stability of its updates, and benefit prediction is performed before business actions are executed to achieve conditional execution. This overcomes the shortcomings of existing technologies, such as data fusion difficulties, inaccurate profiles, and high decision-making risks, improving the accuracy, stability, and automated decision-making level of customer feature analysis.

[0007] The present invention adopts the following technical solution.

[0008] Firstly, a method for analyzing customer electricity consumption characteristics based on neural networks includes: Step 1: Obtain the customer's unstructured event data and time-series electricity consumption data; Step 2: Construct an electricity consumption graph structure to establish a directed association from unstructured event data to time-series electricity consumption data after the event occurs; use the time difference between unstructured event data and time-series electricity consumption data as an influencing factor, and process the electricity consumption graph structure using a graph neural network based on the influencing factor to generate a fused feature vector; Step 3: Using a hierarchical parsing neural network architecture, based on time-series electricity consumption data and fused feature vectors, a deep behavioral representation is generated that simultaneously represents short-term micro-modes and long-term macro-behavior. Step 4: Generate a dynamic customer profile based on deep behavioral representation and update it using a state smoothing algorithm; before automatically executing business actions based on the dynamic customer profile, initiate a simulation decision-making process that includes a virtual control group to estimate the benefits of the business actions; when the estimated benefits meet the preset conditions, execute the business actions.

[0009] Preferably, the unstructured event data and time-series electricity consumption data specifically include: The unstructured event data includes: power service work orders, customer complaint records, electricity inspection reports, and equipment inspection records; The time-series electricity consumption data includes: active power sequence, reactive power sequence, voltage sequence, current sequence, and harmonic data.

[0010] Preferably, the power consumption diagram structure specifically includes: At least one event node has an event feature vector generated based on the unstructured event data and using a text semantic model; At least one time-series data node has a time-series feature vector generated based on the time-series electricity consumption data and using a sequence feature extraction model.

[0011] Preferably, the power consumption diagram structure is processed using a graph neural network based on the influencing factors to generate a fused feature vector, specifically including: A graph attention network mechanism is used to calculate the original attention weights, which are used to represent the basic information transmission strength between the event node and the time-series data node. Based on the aforementioned influencing factors, and using an exponential decay function, the time decay component is calculated; The original attention weights are multiplied by the time decay component to generate the final attention weights; Using the final attention weight, the event feature vectors of the event nodes are weighted and aggregated to update the time-series feature vectors of the time-series data nodes, thereby generating the fused feature vector.

[0012] Preferably, in step 3, the hierarchical analysis of the neural network architecture specifically includes: A convolutional neural network layer is used to take the time-series electricity consumption data as input, and perform sliding convolution operation through the convolution kernel in the convolutional neural network module to extract the short-term micro-mode and generate a micro-mode feature sequence. A recurrent neural network layer is used to perform time-series dependency modeling on the micro-mode feature sequence through a built-in gated recurrent mechanism, parse out the long-period macro-behavior, and generate macro-behavior feature vectors. The feature combination layer is used to concatenate the macroscopic behavioral feature vector with the fused feature vector to generate a combined feature vector; A fully connected network layer is used to receive the combined feature vector and generate the deep behavioral representation.

[0013] Preferably, in step 4, a dynamic customer profile is generated based on deep behavioral representation and updated using a state smoothing algorithm, specifically including: A multi-task learning neural network is used to take the deep behavioral representation as input and generate a pattern probability distribution representing the customer's current behavioral pattern and a quantitative score representing the customer's feature dimension. The pattern probability distribution is combined with the pattern probability distribution of the dynamic customer profile at the previous time step using the Bayesian smoothing method to generate the updated pattern probability distribution. An exponential moving average algorithm is used to perform a weighted average of the quantitative score and the quantitative score of the dynamic customer profile at the previous time step to generate the updated quantitative score.

[0014] Preferably, before automatically executing business actions based on dynamic customer profiles, a simulation decision-making process including a virtual control group is initiated to predict the benefits of the business actions, specifically including: Using a propensity score matching algorithm, based on the quantitative score and historical electricity consumption data as covariates, at least one virtual control user is matched in the historical database for the current customer who meets the triggering conditions. Two scenarios are set up for benefit prediction: a first scenario in which the current customer performs the business action; and a second scenario in which the virtual control user does not perform the business action. The predicted benefits of the first scenario are compared with the predicted benefits of the second scenario to generate the estimated benefit value.

[0015] Secondly, a customer electricity consumption characteristic analysis system based on a neural network, wherein running the aforementioned customer electricity consumption characteristic analysis system based on a neural network includes: The data acquisition module acquires unstructured event data and time-series electricity consumption data from customers; The data fusion module constructs an electricity consumption map structure to establish a directed correlation between unstructured event data and time-series electricity consumption data after the event occurs. The time difference between unstructured event data and time-series electricity consumption data is used as an influencing factor. Based on the influencing factor, a graph neural network is used to process the electricity consumption map structure to generate a fusion feature vector. The feature extraction module adopts a hierarchical parsing neural network architecture, based on time-series electricity consumption data and fused feature vectors, to generate deep behavioral representations that simultaneously represent short-term micro-modes and long-term macro-behaviors. The profiling and decision-making module generates dynamic customer profiles based on deep behavioral representations and updates them using a state smoothing algorithm. Before automatically executing business actions based on the dynamic customer profiles, it initiates a simulation decision-making process that includes a virtual control group to estimate the benefits of the business actions. When the estimated benefits meet preset conditions, the business actions are executed.

[0016] Thirdly, a terminal includes a processor and a storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of a neural network-based customer electricity consumption characteristic analysis method.

[0017] Fourthly, a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of a neural network-based method for analyzing customer electricity consumption characteristics.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention constructs a power consumption graph structure that correlates unstructured events with subsequent time-series data in a directed manner, and uses a graph neural network based on the time difference influence factor for processing. This enables a deep and quantifiable fusion of the semantic information of unstructured data such as work order texts and inspection records with high-dimensional time-series power consumption data. This overcomes the shortcomings of data modality fragmentation, one-sided information utilization, and inability to establish cross-modal causal relationships.

[0019] (2) The hierarchical parsing neural network architecture adopted in this invention can parse time series data layer by layer, first extracting short-term micro-patterns, and then parsing long-term macro-behavior from them. This parsing mechanism from micro to macro, compared with the coarse-grained traditional model, can more accurately and deeply understand complex electricity consumption behavior, and solves the problems of inaccurate dynamic feature extraction and inability to capture deep patterns.

[0020] (3) In view of the defects of delayed and easily mutated portrait updates, the present invention introduces a state smoothing algorithm to ensure that the probability distribution or quantitative score of dynamic portraits is continuously and smoothly evolved when they are updated at high frequency. It can effectively filter data noise, avoid portrait jitter and misjudgment caused by instantaneous behavioral fluctuations, and improve the credibility and usability of portrait results in actual business.

[0021] (4) This invention addresses the pain points of disconnect between analysis and business operations and high decision-making risks by introducing a simulation decision-making process that includes a virtual control group, allowing for scenario-based simulations before business actions are executed. This achieves a complete automated closed loop from data analysis to intelligent decision-making and then to safe execution, overcoming the shortcomings of traditional solutions that blindly push out solutions and have uncontrollable risks, and transforming data value into reliable business benefits. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a customer electricity consumption characteristic analysis method based on neural networks in this invention. Figure 2 This is a schematic diagram of the structure of a customer electricity consumption characteristic analysis system based on neural networks in this invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0024] Example 1: like Figure 1 As shown, this invention provides a method for analyzing customer electricity consumption characteristics based on neural networks, specifically including: Step 1: Obtain the customer's unstructured event data and time-series electricity consumption data; When performing this data acquisition step, data from multiple heterogeneous business systems within the enterprise can be accessed and integrated through application programming interfaces (APIs), direct database connections, or message queues.

[0025] The unstructured event data and time-series electricity consumption data specifically include: The unstructured event data consists of discrete records representing specific business activities or customer states, primarily in free text format. During acquisition, the entire content of the free text is extracted and bound to the corresponding customer ID and event occurrence time. Specifically, this may include: Electricity service work orders, with text such as "User applies for installation of distributed photovoltaic" or "User applies for package change".

[0026] Customer complaint records, texts such as "Users reported that the air conditioner tripped around 3 pm" and "Frequent power outages".

[0027] Electricity inspection reports, texts such as "User found to have illegally connected electrical wires" and "Meter seal abnormality".

[0028] And equipment inspection records, such as "abnormal noise from the transformer" or "signs of overheating at the joints".

[0029] The time-series electricity consumption data is continuous measurement data characterizing user electricity consumption behavior, collected at a fixed sampling frequency using customer ID and timestamp as the primary key. In this embodiment, the sampling frequency is once every 15 minutes, meaning that each customer generates 96 data points per day for each indicator. Specifically, it may include: Active power series is used to characterize the overall electricity consumption level and load shape of users.

[0030] Reactive power sequences are used to analyze the nature of user loads (such as inductive or capacitive) and to determine the usage of large equipment (such as air conditioners and motors). Voltage sequences are used to analyze power quality and power supply stability. Current sequence; In addition, harmonic data is used to identify the connection and usage of nonlinear loads (such as charging piles, variable frequency air conditioners, and switching power supplies).

[0031] By performing the above acquisition steps, the present invention provides a comprehensive raw dataset for subsequent analysis steps and transmits the acquired data to the next step for processing.

[0032] Step 2: Construct an electricity consumption graph structure to establish a directed correlation from unstructured event data to time-series electricity consumption data after the event occurs; use the time difference between unstructured event data and time-series electricity consumption data as an influencing factor, and use a graph neural network to process the electricity consumption graph structure based on the influencing factor to generate a fused feature vector.

[0033] The power consumption diagram structure specifically includes: At least one event node has an event feature vector generated based on the unstructured event data and using a text semantic model; At least one time-series data node has a time-series feature vector generated based on the time-series electricity consumption data and using a sequence feature extraction model.

[0034] Specifically, for each piece of unstructured event data acquired, an event node is created in a graph database (such as Neo4j). Each event node contains at least the following attributes: event ID, customer ID, event occurrence time, and original text content. The original text content of the event node is then fed into a pre-trained text semantic model, such as the Sentence-BERT model, and outputs a fixed-length vector as the event feature vector.

[0035] For time-series electricity consumption data, aggregation is performed at a preset time granularity. In this embodiment, a time-series data node is created for each day within the analysis period for this customer, containing at least the customer ID and date attribute, and associated with the corresponding time-series electricity consumption data for that day. A pre-trained one-dimensional convolutional autoencoder is used as the sequence feature extraction model, outputting a low-dimensional (64-dimensional) time-series feature vector that represents the core waveform features of the daily electricity load curve.

[0036] Automatically query all event nodes and match them with all time-series data nodes that occurred after the event. Establish a directed edge between each event node and each matching time-series data node.

[0037] By performing the above steps, an electrical diagram structure was constructed.

[0038] Based on the influencing factors, a graph neural network is used to process the electricity consumption diagram structure to generate a fused feature vector, specifically including: The original attention weights are calculated using a graph attention network mechanism. , which represents the basic information transmission strength between the event node j and the time-series data node i; this weight is calculated by the graph attention network through its attention mechanism (e.g., a small fully connected network and the LeakyReLU activation function) based on the feature vectors (event feature vector and time-series feature vector) of the two connected nodes.

[0039] Based on the aforementioned influencing factors, and using an exponential decay function, the time decay component is calculated.

[0040] The influencing factor is the time difference between event node j and time series data node i. For example, if the event node date is October 10th and the time series data node date is October 13th, then... Therefore, the exponential decay function Represented as:

[0041] in, The decay rate hyperparameter is positive, and e is the base of the natural logarithm.

[0042] The original attention weights are multiplied by the time decay component to generate the final attention weights. , is represented as:

[0043] Using the final attention weight, the event feature vectors of the event nodes are weighted and aggregated to update the temporal feature vectors of the temporal data nodes, generating the fused feature vector, as follows:

[0044] in, Let i be the set of all adjacent event nodes of time-series data node i. For event feature vectors, It is a learnable weight matrix. This is the aggregated neighbor feature vector. It is obtained by concatenation or summation. With node i's own temporal feature vector Combined, a fused feature vector is generated.

[0045] After all layers of the graph neural network (e.g., a 3-layer graph attention network) have been executed, the final fused feature vector is output, which incorporates the semantic information of all associated text events as they decay over time.

[0046] Step 3: Using a hierarchical parsing neural network architecture, based on time-series electricity consumption data and fused feature vectors, a deep behavioral representation is generated that simultaneously represents short-term micro-modes and long-term macro-behavior. Step 3 involves a hierarchical analysis of the neural network architecture, specifically including: A convolutional neural network layer is used to take the time-series electricity consumption data as input, and perform sliding convolution operation through the convolution kernel in the convolutional neural network module to extract the short-term micro-mode and generate a micro-mode feature sequence. The recurrent neural network layer is used to perform time-series dependency modeling on the micro-mode feature sequence through a built-in gated recurrent mechanism, parse out the long-period macro-behavior, and generate macro-behavior feature vectors.

[0047] The feature combination layer is used to concatenate the macroscopic behavioral feature vector with the fused feature vector to generate a combined feature vector; A fully connected network layer is used to receive the combined feature vector and generate the deep behavioral representation.

[0048] Specifically, in this embodiment, the convolutional neural network layer can be a one-dimensional convolutional neural network (1D-CNN). To make the extracted features interpretable, the 1D-CNN can be pre-trained on a synthetic load database. This database is generated by randomly combining the electrical signatures of known single appliances (such as refrigerators and air conditioners). Through a pre-training task aimed at minimizing the feature correlation between different convolutional kernels, different convolutional kernels can be specialized. For example, one set of convolutional kernels is specifically used to identify short-term pulse features (such as compressor startup), and another set of convolutional kernels is specifically used to identify stable climbing features (such as charging pile startup). The 1D-CNN layer performs a sliding convolution operation on the input 96-point time-series data. Its convolutional kernels sweep across the entire time series, capturing these short-term micro-patterns and outputting a micro-pattern feature sequence (i.e., the CNN feature map sequence). Each time step of this sequence represents the activation intensity of the micro-pattern detected at the corresponding time.

[0049] The recurrent neural network layers can be either Long Short-Term Memory (LSTM) networks or Gated Recurrent Unit (GRU) networks. LSTM or GRU networks receive sequences of micro-pattern features, and their internal gating mechanisms (such as forget gates and input gates) are trained to learn the long-term temporal dependencies and combinatorial grammar of these micro-patterns. For example, it can learn that "(micro-pattern A) occurs periodically + (micro-pattern B) occurs during nighttime lows = (macro-behavior) electric vehicle owners with nighttime charging habits." The output is a macro-behavioral feature vector (e.g., the last hidden state vector of the LSTM network). This characterizes the long-term patterns and behavioral patterns inherent in the entire time series.

[0050] The feature combination layer concatenates two input vectors along their feature dimensions. For example, if the macroscopic behavioral feature vector is 128-dimensional and the fused feature vector is 256-dimensional, the concatenation operation will generate a 384-dimensional combined feature vector.

[0051] The fully connected network layer is implemented as a multilayer perceptron (MLP) (e.g., containing two fully connected layers and a ReLU activation function). The combined feature vector flows through this MLP, undergoing a final nonlinear transformation and feature space projection. The final output is a fixed-dimensional (e.g., 512-dimensional) vector, i.e., a deep behavioral representation. This representation is the sole input for subsequent generation of dynamic profiles, simultaneously encoding micro-patterns, macro-level regularities, and the influence of external events in unstructured text within the time-series data.

[0052] Step 4: Generate a dynamic customer profile based on deep behavioral representation and update it using a state smoothing algorithm; before automatically executing business actions based on the dynamic customer profile, initiate a simulation decision-making process that includes a virtual control group to estimate the benefits of the business actions; when the estimated benefits meet the preset conditions, execute the business actions.

[0053] In step 4, a dynamic customer profile is generated based on deep behavioral representation and updated using a state smoothing algorithm, specifically including: A multi-task learning neural network is used to take the deep behavioral representation as input and generate a pattern probability distribution representing the customer's current behavioral pattern and a quantitative score representing the customer's feature dimension. The pattern probability distribution is combined with the pattern probability distribution of the dynamic customer profile at the previous time step using the Bayesian smoothing method to generate the updated pattern probability distribution. An exponential moving average algorithm is used to perform a weighted average of the quantitative score and the quantitative score of the dynamic customer profile at the previous time step to generate the updated quantitative score.

[0054] Specifically, the multi-task learning neural network can be an MLP that receives the deep behavioral representation from step 3 as input and has two parallel output heads with different activation functions to perform different tasks simultaneously: Output Header A (Classification Header): This output head consists of a fully connected layer and a Softmax activation function. Its output is an N-dimensional pattern probability distribution vector (e.g., [0.75, 0.15, 0.10]), where N is the total number of predefined customer behavior patterns (e.g., commuting to recharge, working from home).

[0055] Output Head B (Regression Head): This output head consists of one or more fully connected layers and a sigmoid activation function. Its output is a K-dimensional quantized score vector (e.g., [0.85, 0.30]), where K is the total number of predefined feature dimensions (e.g., price sensitivity, environmental preference).

[0056] To prevent sudden changes in profile patterns due to noise in daily data, two smoothing algorithms are used: When the system obtains the new mode probability distribution through output header A at time t. At that time, the new distribution is treated as the likelihood, and the old distribution in the dynamic customer profile stored at time t-1 is used. The prior is considered, and the posterior, i.e., the updated image distribution at time t, is calculated through Bayesian inference (simplified in this embodiment to element-wise multiplication followed by normalization). :

[0057] in, The sign is proportional, indicating It is proportional to the product of the two terms on the right.

[0058] At time t, the system obtains a new quantized score through output header B. At that time, the exponential moving average (EMA) algorithm is used to compare it with the old score stored at time t-1. Perform a weighted average:

[0059] in, It is a preset smoothing coefficient (e.g., 0.1).

[0060] Before automatically executing business actions based on dynamic customer profiles (i.e., generated pattern probability distributions and quantitative scores), a simulation decision-making process including a virtual control group is initiated to predict the benefits of the business actions, specifically including: Using a propensity score matching algorithm, based on the quantitative score and historical electricity consumption data as covariates, at least one virtual control user is matched in the historical database for the current customer who meets the triggering conditions. Two scenarios are set up for benefit prediction: a first scenario in which the current customer performs the business action; and a second scenario in which the virtual control user does not perform the business action. The predicted benefits of the first scenario are compared with the predicted benefits of the second scenario to generate the estimated benefit value.

[0061] Specifically, firstly, a preset trigger condition can be set when the current customer's price sensitivity quantitative score is greater than 0.8 for three consecutive days.

[0062] Next, if the current customer meets the above triggering conditions, a temporary logistic regression model is trained using quantitative scores, historical electricity consumption data, and other static attributes (industry classification, geographical region, etc.) as covariates. This model is used to predict the probability that any historical user will meet the triggering conditions, i.e., the propensity score. The historical electricity consumption data includes the customer's average daily electricity consumption, peak-valley difference rate, load factor, and other statistical characteristics over the 30 days prior to the trigger.

[0063] Subsequently, the historical database is searched for all users who have never received the stated business action in the past (e.g., have never been recommended this package), and their propensity scores are calculated. Among these users, one or more users whose propensity scores are extremely close to those of the current customer (e.g., within 0.01) are selected as virtual control users.

[0064] Next, using an XGBoost or Transformer time-series forecasting model, the model receives customer characteristics (the covariates mentioned above) and the upcoming business action, and outputs the expected benefits (e.g., the predicted total electricity cost) for the customer over the next N days (e.g., 30 days). Predictions are then calculated for the following two mutually exclusive scenarios: Scenario 1 (Experimental Group): Input: The current customer's covariates, and the business action set to execution (e.g., simulating the application of a new electricity pricing package). Output: The predicted benefits of the first scenario (e.g., predicting the customer's total electricity bill for the next 30 days to be 850 yuan). Scenario 2 (Control Group): Input: Covariates of the virtual reference user, and business actions that are set to non-execution state (e.g., simulating the action of not applying the new electricity price package).

[0065] Output: Predictive benefits of the second scenario (e.g., predicting the total electricity cost for the control group user over the next 30 days to be 1000 yuan). The predicted benefits of the two scenarios are arithmetically compared, and this 150 / 1000 = 15% is the final estimated benefit value. This value is then compared with preset conditions (e.g., benefit > 10%). If the condition is met, the system ultimately executes the business action; if not, no action is taken, thus achieving a low-risk automated decision-making closed loop.

[0066] Example 2: Based on Example 1, such as Figure 2 As shown, a customer electricity consumption characteristic analysis system based on neural networks includes: The data acquisition module acquires unstructured event data and time-series electricity consumption data from customers; The data fusion module constructs an electricity consumption map structure to establish a directed correlation between unstructured event data and time-series electricity consumption data after the event occurs. The time difference between unstructured event data and time-series electricity consumption data is used as an influencing factor. Based on the influencing factor, a graph neural network is used to process the electricity consumption map structure to generate a fusion feature vector. The feature extraction module adopts a hierarchical parsing neural network architecture, based on time-series electricity consumption data and fused feature vectors, to generate deep behavioral representations that simultaneously represent short-term micro-modes and long-term macro-behaviors. The profiling and decision-making module generates dynamic customer profiles based on deep behavioral representations and updates them using a state smoothing algorithm. Before automatically executing business actions based on the dynamic customer profiles, it initiates a simulation decision-making process that includes a virtual control group to estimate the benefits of the business actions. When the estimated benefits meet preset conditions, the business actions are executed.

[0067] Example 3: A terminal, comprising a processor and a storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of a neural network-based customer electricity consumption characteristic analysis method.

[0068] Example 4: A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of a neural network-based method for analyzing customer electricity consumption characteristics.

[0069] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0070] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0071] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0072] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for analyzing customer electricity consumption characteristics based on neural networks, characterized in that, include: Step 1: Obtain the customer's unstructured event data and time-series electricity consumption data; Step 2: Construct an electricity consumption graph structure to establish a directed association from unstructured event data to time-series electricity consumption data after the event occurs; use the time difference between unstructured event data and time-series electricity consumption data as an influencing factor, and process the electricity consumption graph structure using a graph neural network based on the influencing factor to generate a fused feature vector; Step 3: Using a hierarchical parsing neural network architecture, based on time-series electricity consumption data and fused feature vectors, a deep behavioral representation is generated that simultaneously represents short-term micro-modes and long-term macro-behavior. Step 4: Generate a dynamic customer profile based on deep behavioral representation and update it using a state smoothing algorithm; before automatically executing business actions based on the dynamic customer profile, initiate a simulation decision-making process that includes a virtual control group to estimate the benefits of the business actions; when the estimated benefits meet the preset conditions, execute the business actions.

2. The method for analyzing customer electricity consumption characteristics based on neural networks according to claim 1, characterized in that, The unstructured event data and time-series electricity consumption data specifically include: The unstructured event data includes: power service work orders, customer complaint records, electricity inspection reports, and equipment inspection records; The time-series electricity consumption data includes: active power sequence, reactive power sequence, voltage sequence, current sequence, and harmonic data.

3. The customer electricity consumption characteristic analysis method based on neural networks according to claim 2, characterized in that, The power consumption diagram structure specifically includes: At least one event node has an event feature vector generated based on the unstructured event data and using a text semantic model; At least one time-series data node has a time-series feature vector generated based on the time-series electricity consumption data and using a sequence feature extraction model.

4. The customer electricity consumption characteristic analysis method based on neural networks according to claim 3, characterized in that, Based on the influencing factors, a graph neural network is used to process the electricity consumption diagram structure to generate a fused feature vector, specifically including: A graph attention network mechanism is used to calculate the original attention weights, which are used to represent the basic information transmission strength between the event node and the time-series data node. Based on the aforementioned influencing factors, and using an exponential decay function, the time decay component is calculated; The original attention weights are multiplied by the time decay component to generate the final attention weights; Using the final attention weight, the event feature vectors of the event nodes are weighted and aggregated to update the time-series feature vectors of the time-series data nodes, thereby generating the fused feature vector.

5. The method for analyzing customer electricity consumption characteristics based on neural networks according to claim 1, characterized in that, Step 3 involves a hierarchical analysis of the neural network architecture, specifically including: A convolutional neural network layer is used to take the time-series electricity consumption data as input, and perform sliding convolution operation through the convolution kernel in the convolutional neural network module to extract the short-term micro-pattern and generate a micro-pattern feature sequence. A recurrent neural network layer is used to perform time-series dependency modeling on the micro-mode feature sequence through a built-in gated recurrent mechanism, parse out the long-period macro-behavior, and generate macro-behavior feature vectors. The feature combination layer is used to concatenate the macroscopic behavioral feature vector with the fused feature vector to generate a combined feature vector; A fully connected network layer is used to receive the combined feature vector and generate the deep behavioral representation.

6. The customer electricity consumption characteristic analysis method based on neural networks according to claim 1, characterized in that, In step 4, a dynamic customer profile is generated based on deep behavioral representation and updated using a state smoothing algorithm, specifically including: A multi-task learning neural network is used to take the deep behavioral representation as input and generate a pattern probability distribution representing the customer's current behavioral pattern and a quantitative score representing the customer's feature dimension. The pattern probability distribution is combined with the pattern probability distribution of the dynamic customer profile at the previous time step using the Bayesian smoothing method to generate the updated pattern probability distribution. An exponential moving average algorithm is used to perform a weighted average of the quantitative score and the quantitative score of the dynamic customer profile at the previous time step to generate the updated quantitative score.

7. The method for analyzing customer electricity consumption characteristics based on neural networks according to claim 6, characterized in that, Before automatically executing business actions based on dynamic customer profiles, a simulation decision-making process including a virtual control group is initiated to predict the benefits of the business actions, specifically including: Using a propensity score matching algorithm, based on the quantitative score and historical electricity consumption data as covariates, at least one virtual control user is matched in the historical database for the current customer who meets the triggering conditions. Two scenarios are set up for benefit prediction: a first scenario in which the current customer performs the business action; and a second scenario in which the virtual control user does not perform the business action. The predicted benefits of the first scenario are compared with the predicted benefits of the second scenario to generate the estimated benefit value.

8. A customer electricity consumption characteristic analysis system based on a neural network, comprising the customer electricity consumption characteristic analysis system based on a neural network as described in any one of claims 1-7, characterized in that, include: The data acquisition module acquires unstructured event data and time-series electricity consumption data from customers; The data fusion module constructs an electricity consumption map structure to establish a directed correlation between unstructured event data and time-series electricity consumption data after the event occurs. The time difference between unstructured event data and time-series electricity consumption data is used as an influencing factor. Based on the influencing factor, a graph neural network is used to process the electricity consumption map structure to generate a fusion feature vector. The feature extraction module adopts a hierarchical parsing neural network architecture, based on time-series electricity consumption data and fused feature vectors, to generate deep behavioral representations that simultaneously represent short-term micro-modes and long-term macro-behaviors. The profiling and decision-making module generates dynamic customer profiles based on deep behavioral representations and updates them using a state smoothing algorithm. Before automatically executing business actions based on the dynamic customer profiles, it initiates a simulation decision-making process that includes a virtual control group to estimate the benefits of the business actions. When the estimated benefits meet preset conditions, the business actions are executed.

9. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-7.

Images