Power application oriented data prediction method and system

Abstract

The invention discloses a power application oriented data prediction method and system, and the method comprises the following steps: 1, integrating power customer appeal and feedback data in work order information of a plurality of channels, and obtaining integrated information; step 2, based on the integrated information, establishing a tag attribute management system, manually adding new feature tags, and updating a data feature library on time; 3, considering the high dimension and nonlinearity of the power customer behavior data from the updated data feature library, performing dimension reduction on the updated data features, projecting the data features to a Riemannian manifold, and constructing a power customer behavior prediction model; and 4, on the basis of the constructed prediction model, predicting the power demand of the power customer in a period of time in the future, regularly feeding back a power customer demand processing result and latest customer data to the prediction model, adjusting and optimizing parameters and a structure of the prediction model through feedback data, and performing power application-oriented data prediction. And the prediction accuracy and reliability are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power data processing, and more specifically, relates to a power application-oriented data prediction method and system. Background Art

[0002] Early power data forecasting technology relied on statistical methods and models, such as linear regression analysis and time series analysis. These methods mainly make predictions through statistical analysis and trend fitting of historical data. However, due to the lack of large-scale data and high-performance computing resources, the prediction accuracy and stability of these methods in practical applications are limited. With the improvement of computing power and the development of data acquisition technology, machine learning technology has begun to be applied to the field of power data forecasting. For example, machine learning algorithms such as decision trees, random forests, neural networks, and support vector machines have gradually become mainstream. They can better handle complex data relationships and large-scale data, and improve the accuracy and efficiency of predictions. Carry out joint prevention and control of service risks, move the risk prevention threshold forward, and achieve "key time periods, key units, key customers, and key employees" "controllable, under control, and controllable", and predict in advance possible power supply service risk problems including abnormal electricity charges, abnormal electricity consumption, abnormal business handling, etc., and timely pre-control key work and time periods, key units and equipment, key customers, and key employees. It is particularly important.

[0003] Although relevant prediction technologies are constantly improving, there are still many technical defects: 1. The value of customer demand data is not fully mined. It is not sufficient to mine the value of deep-level relationship data such as customer work order demands and load, collection, voltage and other data in other systems. This has led to some potential abnormal data or risk events not being discovered in a timely manner. 2. There is currently no universal and comprehensive label feature system for customer demand changes, concerns, etc. If customer data features are not fully extracted and important data features that affect customer perception or data anomalies are not fully provided, the prediction model will not be able to capture the diversity and complexity of customer behavior or will have deviations in the understanding of customer behavior, thereby reducing the accuracy of the model's prediction results. 3. Insufficient prediction accuracy. Although existing technologies have adopted machine learning and deep learning methods, the prediction accuracy still needs to be improved. This is mainly because the existing prediction models do not have a deep enough understanding of high-dimensional complex behavior patterns. Summary of the invention

[0004] In order to solve the deficiencies in the prior art, the present invention provides a power application-oriented data prediction method and system.

[0005] The present invention adopts the following technical solution.

[0006] A first aspect of the present invention provides a power application-oriented data prediction method, comprising the following steps:

[0007] Step 1: Integrate the power customer demands and feedback data in the work order information of multiple channels to obtain integrated information;

[0008] Step 2: Based on the integrated information, establish a tag attribute management system, manually add new feature tags, and update the data feature library on time;

[0009] Step 3, from the updated data feature library, taking into account the high dimensionality and nonlinearity of the power customer behavior data, the updated data features are reduced in dimension and projected onto the Riemann manifold to construct a power customer behavior prediction model;

[0010] Step 4: Based on the constructed prediction model, predict the electricity demand of power customers in the future, and regularly feed back the results of customer demand processing and the latest customer data to the prediction model. Adjust and optimize the parameters and structure of the prediction model through the feedback data to perform data prediction oriented to power application.

[0011] Preferably, in step 1, the work order information includes 95598 work orders, 12398 work orders, public opinion work orders and online State Grid; customer information includes business handling information and interaction information of customers in various power supply business channels, archival data including the customer's substation area and line conditions, as well as customer power outage conditions and voltage value information;

[0012] Perform quality checks and cleansing on the extracted data information, including processing missing values, erroneous values, and duplicates; use data fusion to integrate information from various data sources, standardize and uniformly encode data, and establish a unified data model based on business needs; use data warehouses suitable for large data volume storage and query, including Hadoop, Spark, and cloud databases.

[0013] Preferably, in step 2, the establishment of a label attribute management system includes applying natural language processing (NLP) to analyze text content, automatically adding corresponding feature labels to the request work order through keyword extraction and sentiment analysis; creating a dynamic data feature library to provide data features to the prediction model, predict customer behavior and predict data anomalies.

[0014] Preferably, in step 3, constructing the power customer behavior prediction model specifically includes the following steps:

[0015] Step 3.1, define the data manifold M, which is used to represent the customer behavior status. Customer behavior includes multiple dimensions such as demand frequency, demand type, satisfaction score, and load data. All dimensions together constitute a data point (x∈M);

[0016] The Riemannian metric g(x) is defined on the manifold M, which assigns the inner product to all tangent vector spaces TxM on the manifold and is used to measure length and angle;

[0017] Step 3.2, based on the defined data manifold, learn the low-dimensional manifold structure of the actual distribution of data points, and use isometric feature mapping Isomap, local linear embedding and t-SNE algorithm to discover the essential structure of the data;

[0018] Step 3.3, using geometric tools to analyze and model customer behavior on the learned low-dimensional manifold, including: using the Riemann metric (g) to measure the shortest path and curve length between two points to reflect the changing trend of customer behavior;

[0019] Step 3.4, after defining the Riemann metric on the manifold M, use the Riemannian statistics method to analyze the data, modeling through the Riemann mean and Riemann midpoint, and statistically analyzing the central tendency of customer behavior;

[0020] Step 3.5, establish a prediction model, regard the time series data as a set of points defined on a Riemann manifold, and establish a time series prediction model on a Riemann manifold;

[0021] Assign an inner product g to each data point x on the Riemann manifold M x Defined on the tangent space TxM, so that (g) changes continuously at each point;

[0022] Set the time series {t i} i =1 N , where each time point t i Corresponding to a point x on the manifold M i , we need to predict the future time point t n+1 The corresponding point x n+1 ;

[0023] Step 3.6, prediction and optimization, training the prediction model on the Riemann manifold with historical data and optimizing the model parameters;

[0024] Step 3.7, verify and deploy the model, verify the model's predictive ability on a test set that maintains temporal consistency, fine-tune and retrain the model to adapt to new data, and integrate it into the power company's data processing flow through an API.

[0025] Preferably, in step 3.1, g(x) is written as a matrix representation. For any two tangent vectors (u,v∈TxM) on the manifold M, the Riemannian metric is used to define the inner product of any two tangent vectors:

[0026] g x (u,v)=<u,G(x)v>

[0027] Where:

[0028] G(x) is a positive definite symmetric matrix representing the Riemannian metric at the point (x).

[0029] Preferably, in step 3.5, for the electric power customer behavior data being linear structure data, a vector autoregression model is constructed:

[0030] Set the time series data point x i All of them are located on the Riemann manifold M. A vector autoregression model is established for prediction analysis. The model is expressed in the form of Euclidean space as follows: Among them A i represents the coefficient matrix, ε t represents the error term;

[0031] Use the tools of Riemannian geometry to generalize the model:

[0032] Let X t =(x t ,x t-1 ,…,x t-p+1 ) is the state vector at time point t, where p represents the order of the vector autoregressive model;

[0033] The Riemann midpoint formula is used on the manifold M to capture the overall characteristics of the data, defined as a point μ that satisfies the following conditions:

[0034]

[0035] Where:

[0036] w i Represents the weight, which is a positive number and sums to 1, d m Represents the distance induced by the Riemannian metric g. By solving the minimization problem, the optimal vector representation μ on the manifold M is obtained.

[0037] Preferably, in step 3.5, for the electricity customer behavior data being nonlinear structure data, a vector autoregression model on a Riemann manifold is constructed:

[0038] Define the "average" initial estimate x on the manifold t|t-1 ≈μ t|t-1 , determine the initial state and benchmark, including the average and median of customer behavior data over a certain period of time in the past;

[0039] For each iteration, calculate the p-time points x t-1 ,…,x t-p The point μ given by the Riemann mean of t|t-1 ; Calculate the Riemann mean at the current time point on the Riemann manifold, and extract the overall characteristics of customer behavior data over the past period of time for subsequent modeling and prediction;

[0040] In μt|t-1 At , the first p points are projected into the tangent space and converted into vectors in the Euclidean space; the complex data structure on the Riemann manifold is converted into the Euclidean vector space for subsequent model building and analysis;

[0041] In this tangent space, the standard vector autoregression model in Euclidean space is used to transform the predicted point x into t|t(1 Map back onto the manifold;

[0042] The model parameters are fitted by minimizing the Riemann distance of points on the manifold.

[0043] Preferably, step 3.6 specifically includes:

[0044] Step 3.6.1, set f:M→R to represent the smooth objective function defined on the Riemann manifold M and the loss function constructed based on the historical electricity consumption data, which is used to quantify the error between the model prediction and the actual electricity consumption data, and find the point x on M that minimizes f(x);

[0045] On a Riemannian manifold, the Riemannian gradient grad f(x) of the objective function f at a point x∈M represents a vector in the tangent space TxM on the manifold M, which is used to indicate the direction in which f increases fastest at x;

[0046] Step 3.6.2, the Riemann gradient grad f(x) is obtained by converting the Euclidean gradient Project onto the tangent space TxM for calculation;

[0047] Step 3.6.3, at point x k Iterative updates are performed at to find the minimum value of f. The update formula is as follows:

[0048] x k+1 =Exp(x k )(-α k grad f(x k ))

[0049] Where:

[0050] Exp(x k ) represents the exponential mapping, which converts the tangent vector -α k grad f(x k ) is mapped back to the manifold M, α k represents the step size of the kth iteration;

[0051] Step 3.6.4, converting between popular geometric mechanisms and parameter spaces through exponential mapping and logarithmic mapping to capture the geometric characteristics of all electricity customer behavior data;

[0052] The exponential mapping Exp(x):TxM→M represents the mapping from the tangent space TxM at point x to the manifold M itself, mapping the tangent vector to a point on the manifold;

[0053] The logarithmic mapping Logx:M→TxM represents the inverse mapping of the exponential mapping, mapping the points on the manifold back to the vectors in the tangent space;

[0054] Step 3.6.5, implement Riemann gradient descent, including calculating the Riemann gradient and choosing an appropriate step size α k , and use the exponential map to update the iteration points.

[0055] Preferably, in step 3.6.2, It represents the gradient of f in Euclidean space, and the formula is as follows:

[0056]

[0057] Where:

[0058] Proj(TxM) represents a projection operation.

[0059] A second aspect of the present invention provides a power application-oriented data prediction system:

[0060] The data collection and cleaning module is used to collect power-related data from different data sources, clean, convert and standardize the data for subsequent analysis and modeling;

[0061] The feature engineering module is used to further process data to extract features, including using domain knowledge to select features, transform and combine features, and handle missing values ​​and outliers;

[0062] The prediction model building module is used to build a prediction model using statistical modeling to predict future power demand, power supply failures, and line losses;

[0063] The model evaluation and optimization module is used to evaluate the performance of the established prediction model, optimize the model, adjust the model parameters, select different algorithms and features, and optimize the model performance based on the evaluation results;

[0064] The real-time monitoring and early warning module is used to monitor data and issue early warnings based on the results generated by the prediction model.

[0065] Compared with the prior art, the beneficial effects of the present invention include at least:

[0066] (1) By integrating customer data, electricity consumption data, demand data, and channel behavior data from different channels, a more comprehensive and detailed data set is provided, so that the prediction model can take into account multiple aspects of customer electricity consumption, including business processing, electricity consumption, electricity charges, and power supply quality, thereby improving the accuracy and reliability of the prediction.

[0067] (2) By combing through the various characteristic elements of the key links of power supply services, a dynamic key feature management system is constructed, and NLP technology is used to automatically analyze and extract the request work orders and perform feature tagging. This effectively avoids the limitation of mining data features only through the work order acceptance content, and explores deeper needs, so that customer behavior analysis can respond to market and business changes in a timely manner.

[0068] (3) The method of projecting data features onto Riemann manifolds is used to process high-dimensional and nonlinear electricity customer behavior data, deeply explore customer behavior habits, gain insights into the individual and common characteristics of customers, provide more in-depth and accurate customer behavior predictions and data anomaly predictions, and reveal the complex structures and patterns in customer behavior data.

[0069] (4) Based on the optimized model, not only can the customer's electricity demand be predicted, but also abnormal electricity consumption behavior and potential problems can be identified, which helps power supply companies to discover and deal with possible problems in advance, provide customers with personalized and better services, avoid potential service risks or customer dissatisfaction, and improve customer satisfaction and service quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 is a flow chart of power application-oriented data prediction provided according to an embodiment of the present invention;

[0071] Figure 2 It is a diagram for constructing an electric power customer prediction model provided in accordance with an embodiment of the present invention. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

[0073] like Figure 1 As shown, embodiment 1 of the present invention provides a data prediction method oriented to power application, comprising the following steps:

[0074] Step 1, data collection and integration: Integrate customer demands and feedback data from work order information from multiple channels such as 95598 work orders, 12398 work orders, public opinion work orders, and online State Grid, as well as customer business handling information and interaction information in various power supply business channels, and archival data in the marketing system including the customer's substation area and line conditions, as well as customer power outage conditions, voltage values ​​and other information. Perform quality checks on the extracted data and clean the data, including processing missing values, error values, duplicates, etc. Use data fusion technology to integrate information from various data sources, standardize and uniformly encode the data, and establish a unified data model based on business needs to ensure data integrity and consistency. Use data warehouse technologies suitable for large data storage and query, such as Hadoop, Spark and cloud databases.

[0075] Step 2: Establishment of label attribute management system

[0076] Apply natural language processing (NLP) technology to analyze text content, and automatically add corresponding feature tags to the request ticket through keyword extraction, sentiment analysis, etc. Create a dynamic data feature system that allows manual addition of new feature tags to ensure that the data feature library is updated regularly or irregularly as the business develops, so as to provide the prediction model with more comprehensive and professional data signs, so that the prediction model can make more in-depth and accurate predictions of customer behavior and data anomalies.

[0077] Step 3: Build a customer behavior prediction model

[0078] Considering the high dimensionality and nonlinearity of electricity customer behavior data, the data features can be projected onto a Riemann manifold. The electricity customer behavior data is modeled as a set of multidimensional data points. An appropriate Riemann metric is defined in this data space and used to analyze the path and form of customer behavior and predict future behavior and possible service risks. The specific construction process is as follows:

[0079] Step 3.1, define the data manifold and Riemannian metric

[0080] Assume that our data set contains the following data feature dimensions: demand frequency (F), demand type (T), satisfaction score (S), and load data (D). These dimensions constitute a point x∈M in a multidimensional space, and a data manifold M is defined to represent this space.

[0081] For each customer state point x, at that point, we can define a Riemann metric G(x). Define G(x) as some inverse weight matrix:

[0082]

[0083] Among them, σF^2, σT^2, σS^2, σD^2 are the variances of the corresponding features, calculated in the neighborhood of point x. This matrix actually encodes the importance of different features. Features with larger variances contribute less in the inner product of this model, reflecting their greater changes in customer behavior status.

[0084] Consider two tangent vectors u and v in TxM (tangent space at point x), which represent the direction or change in the customer behavior state space.

[0085] Using the Riemann metric, the inner product is used to measure the relationship between the two tangent vectors: g x (u,v)=<u,G(x)v> , where G(x) is a positive definite symmetric matrix representing the Riemannian metric at point (x).

[0086] This method can be used to quantify the "distance" between different customer behavior state points, as well as the "speed" or "curvature" of these behaviors changing in a certain direction. This inner product can be used to calculate the shortest path between two points on the manifold, and can also be used to compare changes in different directions on the manifold to better understand the changing patterns of customer behavior. These metrics can be used in subsequent manifold learning steps to optimally embed high-dimensional data points into low-dimensional space, as well as to determine customer behavior trends and patterns in geometric analysis.

[0087] Step 3.2, Manifold Learning of Data Points

[0088] Using the manifold learning algorithm Isomap (Isometric Mapping), high-dimensional data is represented in low dimensions by calculating the geographical distance between data points, that is, based on the shortest path on the manifold. First, the approximate geographical distance between each pair of data points on the manifold is determined, and then multidimensional scaling analysis is used to construct a low-dimensional representation that keeps these distances as similar as possible.

[0089] Step 3.3, Geometric representation of customer behavior

[0090] On the low-dimensional representation obtained by Isomap, suppose we identify a specific behavior pattern, such as an increase in the frequency of complaints related to electricity bills or a long-term abnormal load data of a certain customer. On this manifold, we can use the Riemann metric to calculate the geographical distance between the data points representing this behavior pattern and other behavior states to understand the geometric relationship and potential change trend between this pattern and other customer behavior states.

[0091] Step 3.4, Statistics under the Riemannian metric

[0092] On this manifold space, we can calculate the Riemann mean to find an "average" customer behavior state. This point is the point that minimizes the average distance to all data points under the Riemann metric. By calculating the Riemann mean between different customer groups, we can gain insights into the central tendency of customer behavior and segment customer groups based on this information.

[0093] Step 3.5: Build a prediction model

[0094] The time series data is regarded as a set of points defined on a Riemann manifold, and a time series prediction model on a Riemann manifold is established.

[0095] Assign an inner product g to each data point x on the Riemann manifold M x It is defined on the tangent space TxM so that (g) changes continuously at each point.

[0096] Set the time series {t i} i =1 N , where each time point t i Corresponding to a point x on the manifold M i , we need to predict the future time point t n+1 The corresponding point x n+1 .

[0097] Preferably, when the electricity customer behavior data is linear structure data, a vector autoregression model is constructed:

[0098] Set the time series data point x i They are all located on the Riemann manifold M. A vector autoregression model is established for predictive analysis. The form of the model in Euclidean space is Among them A i is the coefficient matrix, ε t is the error term.

[0099] Use the tools of Riemannian geometry to generalize the model:

[0100] Let X t =(x t ,x t-1 ,…,x t-p+1 ) is the state vector at time point t, where p is the order of the vector autoregression model.

[0101] The Riemann midpoint formula is used on the manifold M to more accurately capture the overall characteristics of the data, defined as a point μ that satisfies the following conditions:

[0102]

[0103] where w i is the weight, which is a positive number and sums to 1, dm is the distance induced by the Riemannian metric g. By solving the above minimization problem, we can obtain the optimal vector representation μ on the manifold M. Solving the minimization problem can determine the tangent vector of the shortest path between two points on the manifold by logarithmic mapping.

[0104] Preferably, for power customer behavior data that is nonlinear structure data, a vector autoregression model on a Riemann manifold is constructed:

[0105] Define the "average" initial estimate x on the manifold t|t-1 ≈μ t|t-1 , determine an initial state or benchmark, such as the average or median of customer behavior data over a certain period of time in the past;

[0106] For each iteration, we calculate the p-time points x t-1 ,…,x i-p The point μ given by the Riemann mean of t|t-1 . The Riemann mean at the current time point is calculated on the Riemann manifold, and the overall characteristics of customer behavior data over the past period of time are extracted to facilitate subsequent modeling and prediction.

[0107] In μ t|t-1 At , we project the first p points into the tangent space and convert them into vectors in Euclidean space. The complex data structure on the Riemann manifold is converted into a more manageable Euclidean vector space to facilitate subsequent model building and analysis.

[0108] In this tangent space, the standard vector autoregression model in Euclidean space is used, and then the predicted point x is transformed into t|t-1 Map back onto the manifold.

[0109] The model parameters are fitted by minimizing the Riemann distance of points on the manifold.

[0110] Step 3.6, prediction and optimization

[0111] Training the prediction model with historical data and using optimization techniques to optimize the model parameters requires special techniques on Riemannian manifolds to achieve this process:

[0112] Step 3.6.1, objective function and Riemann gradient: Let f:M→R be a smooth objective function defined on the Riemann manifold M. It is a loss function constructed based on historical electricity consumption data and is used to quantify the error between the model prediction and the actual electricity consumption data. The goal is to find a point x on M that minimizes f(x).

[0113] On the Riemann manifold, the Riemann gradient grad f(x) of the objective function f at point x∈M is a vector in the tangent space TxM on the manifold M, which indicates the direction in which f increases fastest at x. The Riemann gradient can better fit the intrinsic structure of the data when updating parameters, which is particularly important for predicting customer electricity usage patterns hidden in the complex geometric structure of the data.

[0114] Step 3.6.2, the Riemann gradient grad f(x) is obtained by converting the Euclidean gradient Project it onto the tangent space TxM for calculation.

[0115] Specifically, if is the gradient of f in Euclidean space, then Where Proj(TxM) represents the projection operation. Through the above calculation, it is ensured that the geometric characteristics of the data are preserved while reducing the loss function, which helps to more accurately capture the behavior patterns of power customers.

[0116] Step 3.6.3, iterative update formula: at point x k Iterate and update to find the minimum value of f. The update formula is as follows:

[0117] x k+1 =Exp(x k )(-α k grad f(x k ))

[0118] Among them, Exp(x k ) is an exponential map that transforms the tangent vector -α k grad f(x k ) is mapped back to the manifold M, α k is the step size of the kth iteration. The above steps can better match the customer's electricity consumption pattern and trend.

[0119] Step 3.6.4, exponential mapping and logarithmic mapping: Because the high-dimensional feature structure of electricity customer behavior data is very complex, exponential mapping and logarithmic mapping can be used to convert between popular geometric structures and parameter spaces, so as to fully capture all important geometric features of electricity customer behavior data. The exponential mapping Exp(x):TxM→M is a mapping from the tangent space TxM at point x to the manifold M itself. It maps the tangent vector to a point on the manifold. This process describes the end point of the shortest path along the direction of the tangent vector. The logarithmic mapping Logx:M→TxM is the inverse mapping of the exponential mapping, which maps the point on the manifold back to the vector of the tangent space.

[0120] Step 3.6.5, Implementation of Gradient Descent: When implementing the Riemannian gradient descent method, the key steps include calculating the Riemannian gradient and choosing an appropriate step size α k, and use the exponential map to update the iteration point.

[0121] Step 3.7, Model Validation and Deployment

[0122] Finally, the model's predictive power needs to be verified on a test set that maintains temporal consistency, and the model needs to be fine-tuned or retrained to adapt to new data. It is integrated into the power company's data processing flow through APIs and other means.

[0123] Step 4: Risk warning and real-time monitoring

[0124] Based on the optimized model, customers' electricity demand in the future can be predicted, and abnormal electricity consumption behavior and potential problems can be identified. By issuing early warnings for potential risks in a timely manner, relevant staff can be supported to respond quickly to prevent complaints and spillover risks caused by escalating customer emotions.

[0125] At the same time, the results of various claim processing and the latest customer data are tracked, and regularly fed back to the label attribute system and customer behavior prediction model for calibration and optimization.

[0126] Preferably, the updating in step 2 refers to a dynamic updating of the data feature library, the purpose of which is to provide an up-to-date and comprehensive feature library when building the prediction model;

[0127] Feedback and calibration in step 4 refers to the continuous optimization of the built model. The purpose is to make adjustments based on the latest data in the application of the predictive model to ensure that the model can always maintain high-precision predictive capabilities as time goes by and customer behavior changes.

[0128] 1. Power demand and abnormal power consumption forecast

[0129] According to the changes in customer electricity consumption patterns, the forecast model is updated regularly to improve the accuracy of the forecast. Different customer groups, such as industrial, commercial and residential users, are segmented and their respective electricity demand is forecasted, because different types of customers have different electricity consumption patterns and abilities to respond to market changes. Combining short-term and long-term forecast results provides support for the dispatch and planning of the power grid.

[0130] Monitor customers' electricity consumption data in real time to promptly detect sudden increases or decreases in electricity consumption or other situations that do not conform to historical electricity consumption patterns. Combined with contextual information such as weather, holidays, and appeal events, accurately judge abnormal situations and automatically identify abnormal electricity consumption outside the normal pattern through prediction models.

[0131] 2. Early warning mechanism and feedback adjustment

[0132] Establish a multi-channel early warning mechanism to ensure that early warning information can be quickly conveyed through SMS, email, mobile APP and other methods. Set up different early warning levels, such as low, medium and high, according to the urgency and abnormality of the incident, and define corresponding response measures for different levels. Clarify the responsible persons and backup emergency response teams under different early warning levels to ensure rapid response.

[0133] Example 1: Emergency situations involving casualties, oxygen inhalation, etc. are defined as high risk, and SMS messages are sent to the relevant person in charge of the unit and the substation manager in the area where they are located to properly allocate emergency repair resources and mobile energy storage power supplies.

[0134] Example 2: Based on the voltage data fed back by the collection system, a certain customer has experienced low voltage during peak hours for seven consecutive days, and the household's electricity consumption and electricity charges are normal. If the customer has reported the low voltage problem through internal and external channels, it is defined as medium risk; if the customer has complained about the low voltage problem, once the low voltage situation is monitored, it is defined as medium risk; if the customer has never reported it, it is defined as low risk.

[0135] Record each warning event and response measure in detail, and evaluate the response effect. Based on feedback information and event handling results, continuously optimize and adjust the warning mechanism and response process to improve warning accuracy and response efficiency.

[0136] Through the above detailed implementation steps, accurate prediction of electricity demand and timely identification of abnormal electricity consumption behavior can be effectively achieved. At the same time, a sound early warning mechanism and rapid response operation can be established to ensure the stability of power grid operation and the safety of customers' electricity use.

[0137] Embodiment 2 of the present invention provides a data prediction system oriented to power applications, including: a data collection and cleaning module, a feature engineering module, a prediction model building module, a model evaluation and optimization module, and a real-time monitoring and early warning module;

[0138] The data collection and cleaning module is used to collect power-related data from different data sources, clean, convert and standardize the data for subsequent analysis and modeling;

[0139] The feature engineering module is used to further process data to extract features, including using domain knowledge to select features, transform and combine features, and handle missing values ​​or outliers;

[0140] The prediction model building module is used to build a prediction model using statistical modeling to predict future power demand, power supply failures, and line losses;

[0141] The model evaluation and optimization module is used to evaluate the performance of the established prediction model, optimize the model, adjust the model parameters, select different algorithms or features, and improve the performance of the model based on the evaluation results;

[0142] The real-time monitoring and early warning module is used to monitor data and issue early warnings based on the results generated by the prediction model.

[0143] The beneficial effects of the present invention include at least:

[0144] 1. By integrating customer data, electricity consumption data, demand data, and channel behavior data from different channels, a more comprehensive and detailed data set is provided, so that the prediction model can take into account multiple aspects of customer electricity consumption, including business processing, electricity consumption, electricity charges, and power supply quality, thereby improving the accuracy and reliability of the prediction.

[0145] 2. By combing through various characteristic elements of key links in power supply services, building a dynamic key feature management system, and applying NLP technology to automatically analyze and extract request work orders and perform feature tagging, we can effectively avoid the limitations of mining data features only through work order acceptance content, explore deeper needs, and enable customer behavior analysis to respond to market and business changes in a timely manner.

[0146] 3. The method of projecting data features onto Riemann manifolds is used to process high-dimensional and nonlinear electricity customer behavior data, deeply explore customer behavior habits, gain insights into the individual and common characteristics among customers, provide more in-depth and accurate customer behavior predictions and data anomaly predictions, and reveal the complex structures and patterns in customer behavior data.

[0147] 4. Based on the optimized model, it can not only predict the customer's electricity demand, but also identify abnormal electricity consumption behavior and potential problems, which helps power supply companies to discover and deal with possible problems in advance, provide customers with personalized and better services, avoid potential service risks or customer dissatisfaction, and improve customer satisfaction and service quality.

[0148] The present invention may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present invention.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A data prediction method for power application, characterized in that: The following steps are involved: Step 1: Integrate the power customer demands and feedback data in the work order information of multiple channels to obtain integrated information; Step 2: Based on the integrated information, establish a tag attribute management system, manually add new feature tags, and update the data feature library on time; Step 3, from the updated data feature library, taking into account the high dimensionality and nonlinearity of the power customer behavior data, the updated data features are reduced in dimension and projected onto the Riemann manifold to construct a power customer behavior prediction model; Step 4: Based on the constructed prediction model, predict the electricity demand of power customers in the future, and regularly feed back the results of customer demand processing and the latest customer data to the prediction model. Adjust and optimize the parameters and structure of the prediction model through the feedback data to perform data prediction oriented to power application.

2. The power application-oriented data prediction method according to claim 1, characterized in that: In step 1, the work order information includes 95598 work orders, 12398 work orders, public opinion work orders and online State Grid; customer information includes customer business handling information and interaction information in various power supply business channels, archival data including the customer's substation and line conditions, as well as customer power outage conditions and voltage value information; Perform quality checks and cleansing on the extracted data information, including processing missing values, error values, and duplicates; Use data fusion to integrate information from various data sources, standardize and uniformly encode data, and establish a unified data model based on business needs; Use data warehouses suitable for storing and querying large amounts of data, including Hadoop, Spark, and cloud databases.

3. The power application-oriented data prediction method according to claim 1, characterized in that: In step 2, the establishment of a label attribute management system includes applying natural language processing (NLP) to analyze text content, automatically adding corresponding feature labels to request work orders through keyword extraction and sentiment analysis; creating a dynamic data feature library to provide data features to the prediction model, predict customer behavior and predict data anomalies.

4. The power application-oriented data prediction method according to claim 1, characterized in that: In step 3, building a power customer behavior prediction model specifically includes the following steps: Step 3.1, define the data manifold M, which is used to represent the customer behavior status. Customer behavior includes multiple dimensions such as demand frequency, demand type, satisfaction score, and load data. All dimensions together constitute a data point (x∈M); The Riemannian metric g(x) is defined on the manifold M, which assigns the inner product to all tangent vector spaces TxM on the manifold and is used to measure length and angle; Step 3.2, based on the defined data manifold, learn the low-dimensional manifold structure of the actual distribution of data points, and use isometric feature mapping Isomap, local linear embedding and t-SNE algorithm to discover the essential structure of the data; Step 3.3, using geometric tools to analyze and model customer behavior on the learned low-dimensional manifold, including: using the Riemann metric (g) to measure the shortest path and curve length between two points to reflect the changing trend of customer behavior; Step 3.4, after defining the Riemann metric on the manifold M, use the Riemannian statistics method to analyze the data, modeling through the Riemann mean and Riemann midpoint, and statistically analyzing the central tendency of customer behavior; Step 3.5, establish a prediction model, regard the time series data as a set of points defined on a Riemann manifold, and establish a time series prediction model on a Riemann manifold; Assign an inner product g to each data point x on the Riemann manifold M x Defined on the tangent space TxM, so that (g) changes continuously at each point; Set the time series {t i } i =1 N , where each time point t i Corresponding to a point x on the manifold M i , we need to predict the future time point t n+1 The corresponding point x n+1 ; Step 3.6, prediction and optimization, training the prediction model on the Riemann manifold with historical data and optimizing the model parameters; Step 3.7, verify and deploy the model, verify the model's predictive ability on a test set that maintains temporal consistency, fine-tune and retrain the model to adapt to new data, and integrate it into the power company's data processing flow through an API.

5. The power application-oriented data prediction method according to claim 4, characterized in that: In step 3.1, g(x) is written as a matrix representation. For any two tangent vectors (u,v∈TxM) on the manifold M, the Riemann metric is used to define the inner product of any two tangent vectors: g x (u,v)=<u,G(x)v> Where: G(x) is a positive definite symmetric matrix representing the Riemannian metric at the point (x).

6. The power application-oriented data prediction method according to claim 4, characterized in that: In step 3.5, for the power customer behavior data as linear structure data, a vector autoregression model is constructed: Set the time series data point x i All of them are located on the Riemann manifold M. A vector autoregression model is established for prediction analysis. The model is expressed in the form of Euclidean space as follows: Among them A i represents the coefficient matrix, ε t represents the error term; Use the tools of Riemannian geometry to generalize the model: Let X t =(x t ,x t-1 ,…,x t-p+1 ) is the state vector at time point t, where p represents the order of the vector autoregressive model; The Riemann midpoint formula is used on the manifold M to capture the overall characteristics of the data, defined as a point μ that satisfies the following conditions: Where: w i Represents the weight, which is a positive number and sums to 1, d m Represents the distance induced by the Riemannian metric g. By solving the minimization problem, the optimal vector representation μ on the manifold M is obtained.

7. The power application-oriented data prediction method according to claim 4, characterized in that: In step 3.5, since the electricity customer behavior data is nonlinear structure data, a vector autoregression model on the Riemann manifold is constructed: Define an "average" initial estimate x on the manifold t|t-1 ≈μ t|t-1 , determine the initial state and benchmark, including the average and median of customer behavior data over a certain period of time in the past; For each iteration, calculate the p-time points x t-1 ,…,x t-p The point μ given by the Riemann mean of t|t-1 ; Calculate the Riemann mean at the current time point on the Riemann manifold, and extract the overall characteristics of customer behavior data over the past period of time for subsequent modeling and prediction; In μ t|t-1 At , the first p points are projected into the tangent space and converted into vectors in the Euclidean space; the complex data structure on the Riemann manifold is converted into the Euclidean vector space for subsequent model building and analysis; In this tangent space, the standard vector autoregression model in Euclidean space is used to transform the predicted point x into t|t-1 Map back onto the manifold; The model parameters are fitted by minimizing the Riemann distance of points on the manifold.

8. The power application-oriented data prediction method according to claim 4, characterized in that: Step 3.6 specifically includes: Step 3.6.1, set f:M→R to represent the smooth objective function defined on the Riemann manifold M and the loss function constructed based on the historical electricity consumption data, which is used to quantify the error between the model prediction and the actual electricity consumption data, and find the point x on M that minimizes f(x); On a Riemannian manifold, the Riemannian gradient grad f(x) of the objective function f at a point x∈M represents a vector in the tangent space TxM on the manifold M, which is used to indicate the direction in which f increases fastest at x; Step 3.6.2, the Riemann gradient grad f(x) is obtained by converting the Euclidean gradient Project onto the tangent space TxM for calculation; Step 3.6.3, at point x k Iterative updates are performed at to find the minimum value of f. The update formula is as follows: x k+1 =Exp(x k )(-α k grad f(x k )) Where: Exp(x k ) represents the exponential mapping, which converts the tangent vector -α k grad f(x k ) is mapped back to the manifold M, α k represents the step size of the kth iteration; Step 3.6.4, converting between popular geometric mechanisms and parameter spaces through exponential mapping and logarithmic mapping to capture the geometric characteristics of all electricity customer behavior data; The exponential mapping Exp(x):TxM→M represents the mapping from the tangent space TxM at point x to the manifold M itself, mapping the tangent vector to a point on the manifold; The logarithmic mapping Logx:M→TxM represents the inverse mapping of the exponential mapping, mapping the points on the manifold back to the vectors in the tangent space; Step 3.6.5, implement Riemann gradient descent, including calculating the Riemann gradient and choosing the appropriate step size α k , and use the exponential map to update the iteration points.

9. The power application-oriented data prediction method according to claim 8, characterized in that: In step 3.6.2, It represents the gradient of f in Euclidean space, and the formula is as follows: Where: Proj(TxM) represents a projection operation.

10. A data prediction system oriented to power application, characterized in that: The data collection and cleaning module is used to collect power-related data from different data sources, clean, convert and standardize the data for subsequent analysis and modeling; The feature engineering module is used to further process data to extract features, including using domain knowledge to select features, transform and combine features, and handle missing values ​​and outliers; The prediction model building module is used to build a prediction model using statistical modeling to predict future power demand, power supply failures, and line losses; The model evaluation and optimization module is used to evaluate the performance of the established prediction model, optimize the model, adjust the model parameters, select different algorithms and features, and optimize the model performance based on the evaluation results; The real-time monitoring and early warning module is used to monitor data and issue early warnings based on the results generated by the prediction model.

Images