Power consumer load prediction method and system fusing RPA-AI technology

Abstract

The invention relates to the technical field of power load prediction, and discloses a power consumer load prediction method and system fusing an RPA-AI technology, and the method comprises the steps: carrying out the data collection through RPA, and carrying out the preprocessing; iCEEMDAN decomposition is carried out, and an intrinsic mode function is obtained; constructing a multi-level power load prediction model, and performing short-term power consumer load prediction and long-term power consumer load prediction according to different prediction time limit demands; and setting multi-level power load prediction model evaluation indexes, and performing performance evaluation and optimization on the multi-level power load prediction model. According to the method, the multi-level power load prediction model is constructed, the prediction requirements of different time scales can be met, the prediction precision and efficiency are remarkably improved, and meanwhile the response capacity to power load transaction is enhanced through the real-time monitoring and early warning functions.

Description

Technical Field

[0001] The present invention relates to the technical field of power load forecasting, and specifically to a method and system for power user load forecasting that integrates RPA-AI technology. Background Art

[0002] As the power and energy industry transitions to a clean, low-carbon new energy system, large-scale renewable energy is gradually replacing the dominant position of traditional coal-fired power, leading to significant changes in the industry's existing planning landscape. Consequently, the traditional fossil energy business model is unable to meet the demands of this new scenario. An office system centered around new business areas such as distributed energy planning, green power and green certificate procurement, and power data analysis and forecasting has emerged as the new model. In this context, there is an urgent need to promote the digitalization, automation, and intelligent upgrade of enterprise management models and business processes. In particular, there is an increasing demand for load forecasting capabilities in the power and energy industry that integrate RPA and AI technologies. Specifically, the transition to a new energy system brings with it a vast array of complex data processing requirements, including real-time monitoring, analysis, and forecasting of power loads. These demands require not only efficient data processing capabilities but also intelligent analysis and forecasting capabilities to help enterprises make informed decisions. The introduction of RPA and AI technologies can effectively improve the level of data processing automation and forecasting accuracy, thereby better addressing the various challenges of the new energy system and meeting the high demands of the power and energy industry for load forecasting capabilities.

[0003] Currently, Robotic Process Automation (RPA) technology can automatically execute repetitive and highly standardized business process tasks, and AI algorithms have high predictive accuracy, self-learning and adaptability, and the ability to process large amounts of data. RPA and AI have been applied in various fields and in predictive and early warning functions.

[0004] Among the current invention patents in this field, the system platform is insufficiently capable of processing the massive data generated by large-scale renewable energy access, resulting in data processing bottlenecks; traditional load forecasting methods cannot meet the volatility and uncertainty brought about by the new energy supply system, showing insufficient accuracy and real-time performance; the lack of intelligent analysis and forecasting functions makes it difficult for the existing system to provide scientific decision-making support for power market operations and sales strategies. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problems solved by the present invention are: the system platform is insufficient in processing the massive data generated by large-scale renewable energy access, resulting in a data processing bottleneck; and the traditional load forecasting method cannot meet the volatility and uncertainty brought about by the new energy supply system.

[0007] To solve the above technical problems, the present invention provides the following technical solution: a method for predicting power user load by integrating RPA-AI technology, comprising:

[0008] Use RPA to collect and pre-process data;

[0009] Perform ICEEMDAN decomposition to obtain the intrinsic mode function;

[0010] Construct a multi-level power load forecasting model to perform short-term and long-term power user load forecasting based on different forecast time requirements;

[0011] Set up multi-level power load forecasting model evaluation indicators to evaluate and optimize the performance of the multi-level power load forecasting model.

[0012] As a preferred solution of the power user load forecasting method integrating RPA-AI technology described in the present invention, wherein: using RPA to collect data includes using RPA to retrieve and select features related to load forecasting on a meteorological platform, and simultaneously collecting environmental factors related to load;

[0013] Use RPA to identify and process missing values ​​and outliers, and use RPA to convert data to a unified scale;

[0014] Perform ICEEMDAN decomposition and divide the load data according to its seasonal characteristics; use the ICEEMDAN algorithm to decompose the load data of each season, extract independent components, and remove noise and non-critical information in the data.

[0015] As a preferred solution of the power user load forecasting method integrating RPA-AI technology described in the present invention, the ICEEMDAN decomposition includes: ICEEMDAN improves the original load time series and adds white noise to the initial signal; the addition of white noise and iterative decomposition of ICEEMDAN generate intrinsic mode functions (IMFs), each of which represents the characteristics of the original sequence at different frequencies. The ICEEMDAN decomposition step formula is expressed as follows:

[0016]

[0017] Among them, y(t) represents the original load sequence, y j (t) represents the noise signal, r n(t) represents the nth order residual component and IMF n (t) represents the nth order residual modal component, E n represents the nth-order mode obtained by the EMD algorithm, M(·) represents the envelope calculation, η j (t) represents the addition of the jth white noise, α n-1 represents the standard deviation of the noise removed;

[0018] BiGRU is selected as the neural network model to process the time series information hidden in the power load. The neural network model has two GRU models with opposite directions, which extract the positive and negative bidirectional time series features of the power load and analyze the dependency relationship. The forward GRU calculation formula is expressed as:

[0019]

[0020] Among them, x t represents the input at time t, h t represents the output at time t, r t Represents the reset gate, z t represents the update gate; W xr 、W hr 、W xz 、W hz 、W hn and W xn represents the weight coefficient matrix, b xr 、b hr 、b xz 、b hz and b xn Represents the bias matrix; σ represents the sigmoid function, tanh represents the hyperbolic tangent function, and ο represents the dot product of the two matrices; the output result of ht is obtained through the forward and reverse bidirectional propagation of GRU.

[0021] As a preferred embodiment of the power user load forecasting method integrating RPA-AI technology described in the present invention, the short-term power user load forecasting includes: using RIME algorithm feature learning to perform parameter optimization, and adopting a greedy algorithm to select the features with the most predictive value based on the importance of the features; constructing a BiGRU model, designing the BiGRU network structure, selecting appropriate initial values ​​for the model weights and biases, and using the feature components extracted by the RIME algorithm as input; using gradient descent or its variants to optimize the model parameters; and generating a prediction result based on the input feature components.

[0022] The frost and ice optimization algorithm RIME optimizes the parameters of the electric load forecast model in different regions and seasons. It constructs a soft frost time search strategy and a hard frost puncture mechanism, and introduces a forward greedy selection mechanism. The algorithm particle update formula is expressed as:

[0023]

[0024] r2 <E

[0025]

[0026] in, represents the new position of the particle after the update, where i and j represent the jth particle of the i-th Rime agent; R best,j represents the jth particle of the best frost in the frost population R; r1 is a random number in the range of (-1, 1), which controls the direction of particle movement; cosθ changes with the number of iterations; β is an environmental factor that simulates the influence of the external environment according to the number of iterations; H is the adhesion, which is a random number in the range of (0, 1) and controls the distance between the centers of two rime particles;

[0027] Refined composite multi-scale fuzzy entropy (RCMFE) extracts load data features in load forecasting. When calculating the fuzzy entropy of load data, RCMFE assists in identifying irregularities or complexities in load data, including periodicity, trends, seasonality, and other change patterns of the data, and spatially reconstructs x. The RCMFE formula is expressed as:

[0028]

[0029] in, Indicates that a new time series is obtained, x0(i) represents the average value of m consecutive x(i); dimension m, similarity tolerance r and coarse-grained length p, calculate the coarse-grained sequence of each and

[0030] As a preferred solution of the power user load forecasting method integrating RPA-AI technology described in the present invention, the long-term power user load forecasting includes the ICEEMDAN-CNN-Informer long-time scale power load forecasting method based on parameter soft sharing mechanism;

[0031] Use refined composite multi-scale fuzzy entropy (RCMFE) for data pattern recognition. When calculating the fuzzy entropy of load data, RCMFE identifies irregularities or complexities in the load data, including periodicity, trends, seasonality, and other patterns of change in the data.

[0032] Construct a convolutional neural network for feature extraction, capturing local patterns and trends in time series data through convolutional layers;

[0033] Feature coupling analysis: Use the Person coefficient to analyze the coupling between different features and determine the most critical features for load forecasting;

[0034] The multi-task model is used to learn the coupling strength between load sequences. The MMoE model considers the dependencies between load sequences in different regions through a soft parameter sharing mechanism.

[0035] Build an informer prediction model and input the weighted features into the informer model;

[0036] The role of the refined composite multi-scale fuzzy entropy RCMFE in load forecasting is to extract the characteristics of load data and understand the irregularity and complexity of load data. The RCMFE formula is expressed as:

[0037]

[0038] Reconstruct the space of x, Indicates that a new time series is obtained, xo(i) represents the average value of m consecutive x(i); dimension m, similarity tolerance r and coarse-grained length p, calculate the coarse-grained sequence of each and

[0039] As a preferred solution of the power user load forecasting method integrating RPA-AI technology described in the present invention, the construction of a convolutional neural network for feature extraction includes: the core of the convolutional neural network is the convolution operation. For an m×n convolution kernel, W is an m×n matrix; each weight w in the convolution kernel W can determine the influence of different types of loads on the final prediction result. The weight corresponding to each load is multiplied by the corresponding pixel x in X and then summed. The calculation formula is expressed as:

[0040]

[0041] Construct a multi-task learning model MMoE to learn the correlation and coupling strength between multiple related tasks; the feature vector of the input layer includes historical load, weather, and time. For the kth expert, the input feature vector is g k (x), the output prediction result is f k (g k (x)), for the gating mechanism, a soft attention mechanism is used to weight the prediction results of each expert, and the weight corresponding to the load is w k (x), the final prediction output formula is expressed as:

[0042]

[0043] During the training process, the cross entropy loss function is used to measure the gap between the model's prediction results and the true load value;

[0044] The deep learning model Informer is used to process time-series data, and the self-attention mechanism is used to capture long-term and short-term dependencies in time series data. It accepts multiple different input feature quantities, allowing the model to assign different weights according to the load sequence or task, while considering the long-term trend and short-term volatility of the load sequence, and capturing the seasonality and periodicity of the load sequence.

[0045] As a preferred embodiment of the power user load forecasting method integrating RPA-AI technology described in the present invention, the performance evaluation and optimization includes selecting three indicators, namely, mean absolute percentage error, coefficient of determination, and root mean square error, to evaluate the prediction results in order to accurately evaluate the prediction accuracy of the proposed model and three selected typical models;

[0046] The mean absolute percentage error formula is expressed as:

[0047]

[0048] The coefficient of determination formula is expressed as:

[0049]

[0050] The root mean square error formula is expressed as:

[0051]

[0052] Where n represents the number of test set data; represents the predicted value of the i-th prediction sample; x i Represents the true value of the i-th test sample; It represents the average value of n data; MAPE represents the relative value of the error between the predicted value and the true value, which is a percentage value. When its value is 0%, it is a perfect model; RMSE represents the absolute value of the error between the predicted value and the true value, and its value range is [0,+∞]; R2 represents the determination coefficient, and its value range is [0,1].

[0053] A power user load forecasting system using any method described in the present invention and integrating RPA-AI technology, wherein:

[0054] Power and energy analysis module: uses RPA technology to automatically collect power and energy transaction data and verify, clean, and classify the data;

[0055] Distributed photovoltaic management module: automatically monitors the operating status of the photovoltaic power generation system and collects power generation data; integrates photovoltaic power generation and user power consumption data for billing; predicts future power generation and optimizes photovoltaic system operation;

[0056] Multi-level load forecasting model construction module: Constructs multi-level power load forecasting models based on different forecast time requirements, including traditional time series models and modern machine learning models;

[0057] The model evaluation and optimization module uses evaluation indicators to evaluate the performance of the constructed prediction model and determine its accuracy and reliability under different prediction time limits; based on the evaluation results, the prediction model is adjusted and optimized to continuously improve the prediction effect.

[0058] A computer device comprises: a memory and a processor; the memory stores a computer program, comprising: the steps of implementing any one of the methods of the present invention when the processor executes the computer program.

[0059] A computer-readable storage medium stores a computer program thereon, comprising: steps of implementing any one of the methods of the present invention when the computer program is executed by a processor.

[0060] The beneficial effects of the present invention are as follows: in view of the different changing characteristics of corporate users' electricity consumption in the short, medium and long term, full consideration is given to the influencing conditions such as corporate users' electricity consumption behavior habits, seasonal factors, social factors, etc., and targeted combination of AI algorithms is achieved by constructing a multi-level power load forecasting model. It can meet the forecasting needs of different time scales, significantly improve the forecasting accuracy and efficiency, and at the same time enhance the response capability to power load fluctuations through real-time monitoring and early warning functions. In addition, by integrating RPA digital labor technology, a multifunctional integrated digital office platform has been designed and developed for regular and volatile businesses such as "short-, medium- and long-term electricity transaction volume and price clearing" and "user electricity consumption monitoring", which can aggregate, analyze and accurately predict data to assist corporate decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive effort. Among them:

[0062] Figure 1 This is an overall flow chart of a method for predicting power user load by integrating RPA-AI technology, provided in the first embodiment of the present invention;

[0063] Figure 2 A design diagram of an RPA core workflow processing system for a power user load forecasting method integrating RPA-AI technology, provided in accordance with the second embodiment of the present invention;

[0064] Figure 3A flowchart of the ICEEMDAN-RIME-BiGRU overall coupling model for a power user load forecasting method integrating RPA-AI technology provided in the first embodiment of the present invention;

[0065] Figure 4 This is a schematic diagram of the ICEEMDAN-CNN-Informer algorithm model for a power user load forecasting method integrating RPA-AI technology, provided in the first embodiment of the present invention;

[0066] Figure 5 A diagram showing the steps for implementing a multi-level power load forecasting model for a power user load forecasting method integrating RPA-AI technology, provided in the first embodiment of the present invention;

[0067] Figure 6 This is a graph showing the prediction results of different models in different seasons for a method for predicting power user load by integrating RPA-AI technology, provided in the second embodiment of the present invention. DETAILED DESCRIPTION

[0068] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0069] Example 1, with reference to Figure 1 , as one embodiment of the present invention, provides a method for power user load forecasting integrating RPA-AI technology, including:

[0070] S1: Use RPA to collect data and perform preprocessing.

[0071] Furthermore, the use of RPA for data collection includes using RPA to retrieve and select features related to load forecasting on the meteorological platform, and simultaneously collect environmental factors related to the load.

[0072] Going further, RPA is used to identify and process missing values ​​and outliers, and RPA is used to convert data to a unified scale.

[0073] Furthermore, ICEEMDAN decomposition is performed to divide the load data according to its seasonal characteristics; the ICEEMDAN algorithm is used to decompose the load data of each season, extract independent components, and remove noise and non-critical information in the data.

[0074] Furthermore, the system leverages RPA technology's information collection, workflow processing, and feedback mechanisms to build an underlying computing module. Based on this foundation, the system integrates monitoring, transmission, computing, operation, and application functions to meet the full-cycle electricity consumption analysis needs of corporate power users. By introducing AI algorithms, a digital office platform for electricity consumption analysis has been developed that integrates RPA and AI technologies.

[0075] It should be noted that if Figure 2 The RPA operational management shown here includes the following: Design and Analysis: Users create and edit flowcharts in the process design module. After design is complete, the process design is saved and sent to the process analysis module. The process analysis module loads and analyzes the design file to generate an executable process structure. Analysis and Execution: After analysis is complete, the process execution module receives the analysis results. The execution module schedules and executes tasks based on the analyzed logic, while also handling any runtime exceptions. Execution and Monitoring: During execution, the execution module continuously records logs and provides status feedback to the monitoring tool, allowing users to observe process execution in real time. If problems arise, users can return to the design module to make adjustments. Iterative Optimization: Based on feedback from the execution module, users may return to the design module to adjust or optimize the process, and then return to the analysis module again, forming a continuous iterative process.

[0076] S2: Perform ICEEMDAN decomposition to obtain the intrinsic mode function.

[0077] Furthermore, the ICEEMDAN decomposition includes: ICEEMDAN improves the original load time series and adds white noise to the initial signal; the addition of white noise and the iterative decomposition of ICEEMDAN generate intrinsic mode functions (IMFs), each of which represents the characteristics of the original sequence at different frequencies. The ICEEMDAN decomposition step formula is expressed as follows:

[0078]

[0079] Among them, y(t) represents the original load sequence, y j (t) represents the noise signal, r n (t) represents the nth order residual component and IMF n (t) represents the nth order residual modal component, E n represents the nth-order mode obtained by the EMD algorithm, M(·) represents the envelope calculation, η j (t) represents the addition of the jth white noise, α n-1 represents the standard deviation of the noise removed.

[0080] Time series data is collected using RPA. ICEEMDAN and CNN are used to perform preliminary screening and feature extraction on the load data, deeply exploring the characteristics of various influencing factors. Based on this, the Person coefficient is used to analyze feature coupling. A multi-task model using the MMoE parameter soft sharing mechanism learns the coupling strength between the electric load and related data to more accurately capture the correlation between them.

[0081] Furthermore, BiGRU is selected as the neural network model to process the time series information hidden in the power load. The neural network model has two GRU models in opposite directions, which extract the positive and negative bidirectional time series features of the power load and analyze the dependency relationship. The forward GRU calculation formula is expressed as:

[0082] r t =σ(W xr x t +b xr +W hr h t-1 +b hr )

[0083] z t =σ(W xz x t +b xz +W hz h t-1 +b hz )

[0084] n t =tanh(r t ο(W hn h t-1 +b hn )+W xn x t +b xn )

[0085] h t =z t οh t-1 +(1-z t )οn t

[0086] Among them, x t represents the input at time t, h t represents the output at time t, r t Represents the reset gate, z t represents the update gate; W xr 、W hr 、W xz 、W hz 、W hn and W xn represents the weight coefficient matrix, b xr、b hr 、b xz 、b hz and b xn Represents the bias matrix; σ represents the sigmoid function, tanh represents the hyperbolic tangent function, and ° represents the dot product of the two matrices; the output result of ht is obtained through the forward and reverse bidirectional propagation of GRU.

[0087] It should be noted that MMoE takes into account the dependencies between load sequences in different regions, enhancing its ability to model complex relationships. Finally, the weighted output features are input into the Informer prediction model of electric load, enhancing the model's robustness in dealing with extreme data and data with high volatility, thereby avoiding errors caused by these factors to a certain extent.

[0088] S3: Construct a multi-level power load forecasting model to perform short-term power user load forecasting and long-term power user load forecasting according to different forecast time requirements.

[0089] Furthermore, the short-term power user load forecast is as follows: Figure 3 The method includes using RIME algorithm feature learning to perform parameter optimization, adopting greedy algorithm to select the features with the most predictive value based on the importance of features; building BiGRU model, designing BiGRU network structure, selecting appropriate initial values ​​for the weights and biases of the model, and taking the feature components extracted by RIME algorithm as input; using gradient descent or its variants to optimize model parameters; and generating prediction results based on the input feature components.

[0090] Furthermore, the frost and ice optimization algorithm RIME optimizes the parameters of the electricity load forecast model in different regions and seasons. It constructs a soft frost time search strategy and a hard frost puncture mechanism, and introduces a forward greedy selection mechanism. The algorithm particle update formula is expressed as:

[0091]

[0092] r2 <E

[0093]

[0094] in, represents the new position of the particle after the update, where i and j represent the jth particle of the i-th Rime agent; R best,j represents the jth particle of the best frost in the frost population R; r1 represents a random number in the range of (-1, 1), which controls the direction of particle movement; cosθ changes with the number of iterations; β represents the environmental factor, which simulates the influence of the external environment with the number of iterations; H represents the adhesion, which is a random number in the range of (0, 1) and controls the distance between the centers of two rime particles.

[0095] Furthermore, the refined composite multi-scale fuzzy entropy (RCMFE) extracts load data features in load forecasting. When calculating the fuzzy entropy of load data, RCMFE assists in identifying irregularities or complexities in load data, including periodicity, trend, seasonality, and other change patterns of the data, and spatially reconstructs x. The RCMFE formula is expressed as:

[0096]

[0097] in, Indicates that a new time series is obtained, x0(i) represents the average value of m consecutive x(i); dimension m, similarity tolerance r and coarse-grained length p, calculate the coarse-grained sequence of each and

[0098] Furthermore, the long-term power user load forecasting includes an ICEEMDAN-CNN-Informer long-term power load forecasting method based on parameter soft sharing mechanism, such as Figure 4 shown.

[0099] Furthermore, a refined composite multi-scale fuzzy entropy (RCMFE) is used for data pattern recognition. When calculating the fuzzy entropy of load data, RCMFE identifies irregularities or complexities in the load data, including periodicity, trends, seasonality, and other change patterns of the data.

[0100] Going further, a convolutional neural network is constructed for feature extraction, capturing local patterns and trends in time series data through convolutional layers.

[0101] Furthermore, feature coupling analysis uses the Person coefficient to analyze the coupling between different features and determine the features that are most critical to load forecasting.

[0102] Furthermore, a multi-task model is used to learn the coupling strength between load sequences; the MMoE model considers the dependency between load sequences in different regions through a soft parameter sharing mechanism.

[0103] Furthermore, an Informer prediction model is constructed and the weighted features are input into the Informer model.

[0104] Furthermore, the role of the refined composite multi-scale fuzzy entropy RCMFE in load forecasting is to extract the characteristics of load data and understand the irregularity and complexity of load data. The RCMFE formula is expressed as:

[0105]

[0106] Furthermore, we can reconstruct the space of x. Indicates that a new time series is obtained, xo(i) represents the average value of m consecutive x(i); dimension m, similarity tolerance r and coarse-grained length p, calculate the coarse-grained sequence of each and

[0107] Furthermore, the construction of a convolutional neural network for feature extraction includes the following: the core of the convolutional neural network is the convolution operation. For an m×n convolution kernel, W is an m×n matrix. Each weight w in the convolution kernel W can determine the influence of different types of loads on the final prediction result. The weight corresponding to each load is multiplied by the corresponding pixel x in X and then summed. The calculation formula is expressed as:

[0108]

[0109] Furthermore, a multi-task learning model MMoE is constructed to learn the correlation and coupling strength between multiple related tasks; the feature vector of the input layer includes historical load, weather, and time. For the kth expert, the input feature vector is g k (x), the output prediction result is f k (g k (x)), for the gating mechanism, a soft attention mechanism is used to weight the prediction results of each expert, and the weight corresponding to the load is w k (x), the final prediction output formula is expressed as:

[0110]

[0111]

[0112] Furthermore, during the training process, the cross entropy loss function is used to measure the gap between the model's prediction results and the true load value.

[0113] Furthermore, the deep learning model Informer is used to process time-series data, and the self-attention mechanism is used to capture long-term and short-term dependencies in time series data. It accepts multiple different input feature quantities, allowing the model to assign different weights according to the load sequence or task, while considering the long-term trend and short-term volatility of the load sequence, and capturing the seasonality and periodicity of the load sequence.

[0114] It should be noted that MMoE (Multi-gate Mixture-of-Experts) is a multi-task learning architecture used to learn the correlation and coupling strength between multiple related tasks. It includes multiple "expert" sub-models, each of which is responsible for handling different tasks. In the electric load sequence data, each load sequence can be regarded as a task. The gating network is used to determine the weight of the load, how to combine the output of the experts, and learn the correlation; if different sequences have strong correlation at a certain point in time, the gating mechanism can increase the corresponding weight, and the gating mechanism can be used to learn the coupling strength of the load sequences. There is a strong correlation between the load sequences in different regions, and the gating mechanism will assign higher weights to these tasks to better capture the coupling strength between them.

[0115] S4: Set multi-level power load forecasting model evaluation indicators to perform performance evaluation and optimization on the multi-level power load forecasting model.

[0116] Further, such as Figure 5 As shown in the figure, after building a multi-layered power load forecasting model, the overall design of the integrated office platform system for the power and energy industry based on RPA digital workforce includes the monitoring layer, which is primarily responsible for data collection and preliminary processing. This layer utilizes a variety of devices and technologies, including smart meters, sensors, surveillance cameras, satellite remote sensing, and third-party data sources, to obtain relevant business data for the power and energy industry in real time. Monitoring equipment should be highly accurate, reliable, and adaptable to environmental conditions to ensure data accuracy and timeliness.

[0117] Furthermore, the transport layer is primarily responsible for data transmission and communication. This layer enables efficient data access and interaction through various means, including mobile terminals, fiber-optic private networks, and the carrier internet. The transport protocol uses TCP / IP to ensure stable and secure data transmission. Furthermore, encryption technologies (such as SSL / TLS) can be used to protect data privacy and prevent theft or tampering during transmission.

[0118] Furthermore, the computing layer is primarily responsible for data storage, computation, and analysis. Leveraging the cloud platform's computing, storage, and network resources, this layer combines "two libraries and one platform" (i.e., database, knowledge base, and computing platform) to achieve efficient data processing through intelligent components such as the graphical user interface (GUI) and application programming interface (API). The computing layer also supports RPA process design and management, providing service-oriented support and ensuring flexible and scalable data processing.

[0119] Furthermore, the operations layer is primarily responsible for specific business operations and process automation. This layer includes concrete operations such as data collection, statistical analysis, report generation, and resource downloads. The operations layer connects upward with the application layer to clarify specific task requirements; downward with the computing layer to drive the underlying processing systems of the RPA digital workforce for analysis, calculation, and feedback. The design of the operations layer should prioritize efficiency and accuracy to ensure the smooth operation of business processes.

[0120] Furthermore, the application layer is primarily responsible for implementing specific business functions and automated processes. This layer includes four modules and fifteen sub-functions: power energy trading and data analysis, distributed photovoltaic operations and electricity cost calculation, comprehensive financial management and risk assessment, and enterprise operations management and market development. The design of the application layer should focus on user experience and functional integrity, ensuring that users can easily and quickly access the various services provided by the system.

[0121] Furthermore, the performance evaluation and optimization includes selecting three indicators, namely, mean absolute percentage error, coefficient of determination, and root mean square error, to evaluate the prediction results in order to accurately evaluate the prediction accuracy of the model proposed in this paper and the three typical models selected.

[0122] Furthermore, the mean absolute percentage error formula is expressed as:

[0123]

[0124] Furthermore, the coefficient of determination formula is expressed as:

[0125]

[0126] Furthermore, the root mean square error formula is expressed as:

[0127]

[0128] Where n represents the number of test set data; represents the predicted value of the i-th prediction sample; x i Represents the true value of the i-th test sample; It represents the average value of n data; MAPE represents the relative value of the error between the predicted value and the true value, which is a percentage value. When its value is 0%, it is a perfect model; RMSE represents the absolute value of the error between the predicted value and the true value, and its value range is [0,+∞]; R2 represents the determination coefficient, and its value range is [0,1].

[0129] It should be noted that the Rime Optimization Algorithm (RIME) is a search algorithm that simulates soft and hard frost growth paths under natural conditions. It is primarily used to optimize parameters for electricity load forecasting models in different regions and seasons. By simulating the movement of unfrozen frost particles and the overlapping and intersecting growth of hard frost particles after freezing, the RIME algorithm constructs a soft frost time search strategy and a hard frost penetration mechanism. It also introduces a forward greedy selection mechanism, enabling the algorithm to achieve optimal solution exploration and development in optimization methods.

[0130] On the other hand, this embodiment also provides a power user load forecasting system that integrates RPA-AI technology, which includes:

[0131] Power and energy analysis module: Uses RPA technology to automatically collect power and energy transaction data and verify, clean, and classify the data.

[0132] Distributed photovoltaic management module: automatically monitors the operating status of the photovoltaic power generation system and collects power generation data; integrates photovoltaic power generation and user power consumption data for billing; predicts future power generation and optimizes photovoltaic system operation.

[0133] The multi-level load forecasting model construction module builds a multi-level power load forecasting model based on different forecast time requirements, including traditional time series models and modern machine learning models.

[0134] The model evaluation and optimization module uses evaluation indicators to evaluate the performance of the constructed prediction model and determine its accuracy and reliability under different prediction time limits; based on the evaluation results, the prediction model is adjusted and optimized to continuously improve the prediction effect.

[0135] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0136] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0137] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0138] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0139] Example 2, reference Figure 6 , which is an embodiment of the present invention, provides a method for power user load forecasting that integrates RPA-AI technology. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0140] Using RPA, we selected a regional power grid load dataset for research. The dataset spans 24 months, with a 15-minute sampling interval. It includes load data, date type, meteorological factors, and other data, totaling 70,080 data sets. The dataset is divided into four quarters: spring load from March to May, summer load from June to August, autumn load from September to November, and winter load from December to February. This paper selects the maximum, minimum, and average temperature; weather conditions (categorized as clear, cloudy, hot, rainy, foggy, and cold); air humidity; weekdays, and whether they fall on weekends or holidays as model input features. Missing and abnormal data are filled using linear interpolation and normalized. A 7:2:1 ratio is used for model training, validation, and testing.

[0141] For the training set, ICEEMDAN was used to decompose the raw load data. After repeated experiments, the noise standard deviation of the ICEEMDAN decomposition was ultimately selected to be 0.5, the average number of signal processing cycles was 500, and the maximum number of iterations was 2000. After decomposing the historical load time series for each of the four seasons, nine IMFs and Res components were obtained for different frequency bands. Noise was then filtered out by setting a threshold of 0.1. The remaining useful signal IMFs were then denoised and reconstructed, resulting in five IMFs with empirical modes for different frequency bands. This reduced the non-stationarity of the original data, and no significant modal aliasing was observed in any of the components.

[0142] Using the RIME algorithm, we used load data from each season as input to optimize the hyperparameters of the BiGRU. We set the learning rate lr to the optimization range of [10-5, 0.2] and the maximum number of iterations N to the optimization range of [100, 1500]. We performed a grid search over window lengths of {10, 16, 24, 36, 46, 60} and batch sizes of {16, 32, 64, 128, 256}. For the spring load characteristic component, the algorithm achieved the highest performance when the learning rate lr was 10-2, the number of iterations was 1000, the window length was 36, and the batch size was 64.

[0143] In order to further verify the effectiveness and accuracy of the model proposed in this paper, three typical short time series load forecasting models, BiGRU, ICEEMDAN-BiGRU (I-BiG) and ICEEMDAN-RIME-BiLSTM (IR-BiL), are selected for comparative analysis. The model calculation results are shown in the figure below. Figure 6 As shown in the figure, a day in the middle month of each quarter is randomly selected to show the fit between the load forecast values ​​of each model and the actual power load values ​​in different time periods. The prediction results are shown in Table 1.

[0144] Depend on Figure 6 As can be seen, from a time perspective, the daily load forecast curve values ​​of all models from 10 PM to 4 AM across different seasons are essentially consistent with the actual values. This is because during this period, there is less mobility, less external interference, and less variation in power load. From an overall load perspective, the power load is highest in winter and summer, and lower in spring and autumn. This is because the region experiences significant seasonal differences. Due to the extreme temperatures in winter and summer, maintaining a comfortable temperature requires significant power consumption. From the overall trend of the curve, the daily load curve is at its lowest point from 10 PM to 6 AM. From 8 AM to 2 PM, the daily load curve continues to rise. Due to the low winter temperatures, the winter daily load curve remains at a high level from 2 PM to midnight, a characteristic not seen in spring and autumn. The summer daily load curve reaches its highest value at the hottest time of the day. These models, after learning the load characteristics of different seasons, all reflect these life patterns, demonstrating the accuracy of this paper's predictions based on seasonal differences. Specifically, in different models, the daily load forecast curve of the control group has a large deviation from the actual value, while the overall trend of IR-BiG is very close to the change trend of the actual power load value.

[0145] Table 1. Prediction results of four models in different seasons

[0146]

[0147] Based on the evaluation metrics developed in this paper, the forecast accuracy of different models for different seasons was calculated, as shown in Table 1. As can be seen from the table, due to ICEEMDAN's excellent ability to handle frequency aliasing and pseudo-modal analysis, the RMSE values ​​of the I-BiG model for all four seasons are between 77.46% and 80.85% of those of the BiGRU model. The mean MAPE and R² accuracies are improved by 5.15 and 0.118, respectively. BiLSTM and BiGRU are both types of bidirectional RNNs. The RMSE of the IR-BiG model averages 0.477 times that of the IR-BiL model across all four seasons, while the mean MAPE and R² accuracies improve by 4.025 and 0.081, respectively. This indicates that the BiGRU outperforms the BiLSTM model in load forecasting with seasonal variations. The mean RMSE of the IR-BiG model for all four seasons is between 21.44% and 31.92% of that of the I-BiG model. The proposed IR-BiG algorithm improves the mean MAPE and R² accuracy of BiGRU, I-BiG, and IR-BiL by 15.4%, 10.25%, and 4.02%, and 41.21%, 20.04%, and 9.28%, respectively. This demonstrates the superiority of the proposed IR-BiG method in seasonal short-term load forecasting.

[0148] An ICEEMDAN-CNN-Informer long-time scale electric load forecasting method based on parameter soft sharing mechanism.

[0149] The electric load forecasting model based on the parameter soft sharing mechanism ICEEMDAN-CNN-Informer established in this paper has an input window size of 336×10 and a step size of 1. The load data is decomposed and processed using the ICEEMDAN method. Sequences of components with similar entropy values ​​are reorganized based on fine-grained composite multi-scale fuzzy entropy. A CNN is then used to extract features from the load sequence. The feature extraction layer of the CNN consists of two convolutional layers, two ReLU activation functions, and a fully connected layer. The number of convolution kernels is 10 and 10, and the size is 3×3, which can extract features for each type of load. Secondly, a multi-task model with a MMoE parameter soft-sharing mechanism is used to learn the coupling strength between load sequences. Experiments show that the MMoE model, which includes 10 expert networks and 4 gated networks, performs best. Each expert network is implemented as a single-layer network with 16 hidden layers, which can fully analyze the correlation between different features and predict them separately considering the coupling strength between different features. Finally, the weighted sum of the output features is input into the informer prediction model, thereby achieving accurate medium- and long-term load prediction taking into account coupling uncertainty. The informer model performs grid search on hyperparameters. The encoder contains a three-layer stack and a two-layer decoder. Finally, the model proposed in this patent is compared with the other six algorithms in terms of long-term prediction results to reflect the performance differences of different models in multi-energy microgrid load prediction.

[0150] To verify the accuracy and applicability of the proposed MMoE+CNN+Informer combined forecasting model, we selected two long timescales: 300 hours and 600 hours. We first compared and analyzed the load forecasting results of seven models at the 300-hour scale: CNN, BiGRU, CNN-BiGRU, MMoE-Informer, Informer, MMoE+Informer, and ICEEMDAN-CNN-Informer. In the 300-hour scale, the models using the Informer and BiGRU algorithms showed no significant difference in load forecasting performance over the first 200 hours. However, after 200 hours, the model combining the MMoE and Informer algorithms showed significant improvement. The numerical results for the mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination (R²) for the seven models are shown in Table 2.

[0151] Table 2 Comparison of indicators of different prediction models for 300h sequence

[0152]

[0153] Table 2 shows the performance of CNN, BiGRU, CNN+BiGRU, MMoE+CNN+BiGRU, Informer, MMoE+Informer, and the proposed ICEEMDAN+CNN+Informer model, respectively. It can be seen that the proposed ICEEMDAN+CNN+Informer model has smaller MAPE and RMSE on the 300-hour timescale series, and its R² is closer to 1, which is an improvement over the other models and is relatively superior.

[0154] To ensure objectivity and fairness in the control experiment, the experimental results are averaged after 15 runs, reducing possible accidental errors. This paper also analyzes 600-hour load forecasting. The larger the forecast scale, the more significant the improvement in prediction performance of the model using the MMoE + Informer algorithm. The specific numerical results of the mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination (R²) for the seven models at the 600-hour scale are shown in Table 3.

[0155] Table 3 Comparison of indicators of different prediction models for 600h

[0156]

[0157] Table 3 above shows that the ICEEMDAN+CNN+Informer model proposed in this patent has higher prediction accuracy on the 600h scale. Compared with the MMoE+CNN+BiGRU model, which is also a hybrid model, and the MMoE+Informer model using the Informer algorithm, the RMSE is reduced by 5.05% and 1.05% respectively in power load forecasting; the MAPE values ​​are reduced by 14.66% and 5.79% respectively; and the R2 is closer to 1 than the other six methods. Therefore, the ICEEMDAN+CNN+Informer model proposed in this paper has higher prediction accuracy on the 600h long time scale series compared with the other hybrid models and models that do not use the Informer algorithm.

[0158] Comparing Tables 2 and 3, we can see that for 600-hour timescales, the Informer model achieves higher prediction accuracy than the model without the Informer model. However, for 300-hour timescales, this advantage over the BiGRU model is not significant. However, for 300-hour timescales, the MMoE model's coupling processing and the CNN model's feature preprocessing reduce the MAPE by 5.79% compared to the model without these two algorithms. This shows that the Informer model demonstrates a significant advantage in long-term 600-hour predictions, especially for long series, demonstrating its robustness and improved prediction accuracy. While maintaining superior short-term prediction performance, the Informer model achieves approximately 14.66% and 5.05% lower MAPE and RMSE, respectively, for long-term predictions compared to the algorithm without the Informer model, with the R² approaching 1.

[0159] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for predicting power user load by integrating RPA-AI technology, characterized in that: include: Use RPA to collect data and perform preprocessing; Perform ICEEMDAN decomposition to obtain the intrinsic mode function; Construct a multi-level power load forecasting model to conduct short-term and long-term power user load forecasting according to different forecast time requirements; Set up multi-level evaluation indicators for the power load forecasting model and perform performance evaluation and optimization on the multi-level power load forecasting model.

2. The power user load forecasting method integrating RPA-AI technology as claimed in claim 1 is characterized by: The data collection using RPA includes using RPA to retrieve and select characteristic data related to load forecasting on the meteorological platform, and collecting environmental factor data related to the load; Use RPA to identify and process missing values ​​and outliers, and use RPA to convert data to a uniform scale; ICEEMDAN decomposition is performed to divide the load data according to its seasonal characteristics; the load data of each season is decomposed using the ICEEMDAN algorithm to extract independent components and remove noise and non-critical information from the data.

3. The power user load forecasting method integrating RPA-AI technology as claimed in claim 2 is characterized by: The ICEEMDAN decomposition includes: ICEEMDAN improves the original load time series and adds white noise to the initial signal; adding white noise and iterative decomposition of ICEEMDAN generate intrinsic mode functions IMF, each IMF represents the characteristics of the original sequence at different frequencies, and the decomposition step formula of ICEEMDAN is expressed as: Among them, y(t) represents the original load sequence, y j (t) represents the noise signal, r n (t) represents the nth order residual component and IMF n (t) represents the nth order residual modal component, E n represents the nth-order mode obtained by the EMD algorithm, M(·) represents the envelope calculation, η j (t) represents the addition of the jth white noise, α n-1 represents the standard deviation of noise removal; BiGRU is selected as the neural network model to process the hidden time series information in the power load. The neural network model has two GRU models with opposite directions, which extract the positive and negative bidirectional time series features of the power load and analyze the dependency relationship. The forward GRU calculation formula is expressed as: Among them, x t represents the input at time t, h t represents the output at time t, r t Represents the reset gate, z t represents the update gate; W xr , W hr , W xz , W hz , W hn and W xn represents the weight coefficient matrix, b xr 、b hr 、b xz 、b hz and b xn represents the bias matrix; σ represents the sigmoid function, tanh represents the hyperbolic tangent function, and ° represents the dot product of the two matrices; the output result of ht is obtained through the forward and reverse bidirectional propagation of GRU.

4. The power user load forecasting method integrating RPA-AI technology as claimed in claim 3 is characterized by: The short-term power user load forecasting includes using a short-term load forecasting method based on ICEEMDAN-RIME-BiGRU that takes into account seasonal differences to perform ultra-short-term and short-term forecasts; ultra-short-term is within one day, and short-term is within seven days; Use RIME algorithm feature learning to optimize parameters, and adopt greedy algorithm to select the most predictive features based on the importance of features; Build the BiGRU model, design the BiGRU network structure, select appropriate initial values ​​for the model's weights and biases, and use the feature components extracted by the RIME algorithm as input; Optimize model parameters using gradient descent or its variants; generate predictions based on the input feature components; The frost optimization algorithm RIME uses a soft frost time search strategy and a hard frost puncture mechanism to optimize the parameters of the electric load forecasting model in different seasons under different regional locations. It introduces a positive greedy selection mechanism, and the algorithm particle update formula is expressed as: in, represents the new position of the particle after the update, where i and j represent the jth particle of the i-th Rime agent; R best,j represents the jth particle of the best frost in the frost population R; r1 represents a random number in the range of (-1,1), which controls the direction of particle movement; cosθ changes with the number of iterations; β represents the environmental factor, which simulates the influence of the external environment with the number of iterations; H represents the adhesion, which is a random number in the range of (0,1) and controls the distance between the centers of two rime particles; Refined composite multi-scale fuzzy entropy RCMFE, extracts load data features in load forecasting. When calculating the fuzzy entropy of load data, RCMFE assists in identifying irregularities or complexities in load data, including periodicity, trend, seasonality and other change patterns of data, and spatially reconstructs x. The RCMFE formula is expressed as: in, Indicates that a new time series is obtained, x0(i) represents the average value of m consecutive x(i); dimension m, similarity tolerance r and coarse-grained length p, calculate each coarse-grained sequence and 5. The power user load forecasting method integrating RPA-AI technology as claimed in claim 4 is characterized by: The long-term power user load forecasting includes using the ICEEMDAN-CNN-Informer long-time scale power load forecasting method based on parameter soft sharing mechanism to perform medium-term and long-term power user load forecasting; wherein the medium-term is within half a month, and the long-term is within one month; Use refined composite multi-scale fuzzy entropy RCMFE for data pattern recognition, build a convolutional neural network for feature extraction, and capture local patterns and trends in time series data through convolutional layers; Characteristic coupling analysis, using Person coefficient to analyze the coupling between different characteristics; determine the most critical characteristics for load forecasting; The multi-task model is used to learn the coupling strength between load sequences. The MMoE model considers the dependency between load sequences in different regions through a soft parameter sharing mechanism. Build an informer prediction model and input the weighted features into the informer model; The role of the refined composite multi-scale fuzzy entropy RCMFE in load forecasting is to extract the characteristics of load data and understand the irregularity and complexity of load data. The RCMFE formula is expressed as: Reconstruct x spatially, Indicates that a new time series is obtained, xo(i) represents the average value of m consecutive x(i); dimension m, similarity tolerance r and coarse-grained length p, calculate each coarse-grained sequence and 6. The power user load forecasting method integrating RPA-AI technology as claimed in claim 5 is characterized by: The construction of the convolutional neural network for feature extraction includes: the core of the convolutional neural network is the convolution operation. For a convolution kernel of size m×n, W is an m×n matrix; each weight w in the convolution kernel W can determine the weight corresponding to each load for different types of loads and their influence on the final prediction result, and multiply it with the corresponding pixel x in X and then sum it. The calculation formula is expressed as: Construct a multi-task learning model MMoE to learn the correlation and coupling strength between multiple related tasks; the feature vector of the input layer includes historical load, weather, and time. For the kth expert, the input feature vector is g k (x), the output prediction result is f k (g k (x)), for the gating mechanism, a soft attention mechanism is used to weight the prediction results of each expert, and the weight corresponding to the load is w k (x), the final prediction output formula is expressed as: During the training process, the cross entropy loss function is used to measure the gap between the model's prediction results and the true load value; The deep learning model Informer is used to process time-series data, and the self-attention mechanism is used to capture the long-term and short-term dependencies in time series data. It accepts multiple different input feature quantities, allows the model to assign different weights according to the load sequence or task, considers the long-term trend and short-term volatility of the load sequence, and captures the seasonality and periodicity of the load sequence.

7. The power user load forecasting method integrating RPA-AI technology as claimed in claim 6 is characterized by: The performance evaluation and optimization include, in order to accurately evaluate the prediction accuracy of the model proposed in this paper and the three typical models selected, selecting the mean absolute percentage error, determination coefficient, and root mean square error to evaluate the prediction results; The mean absolute percentage error formula is expressed as: The formula for the coefficient of determination is: The root mean square error formula is expressed as: Where n represents the number of test set data; represents the predicted value of the i-th prediction sample; x i Represents the true value of the i-th test sample; It represents the average value of n data; MAPE represents the relative value of the error between the predicted value and the true value, which is a percentage value. When its value is 0%, it is a perfect model; RMSE represents the absolute value of the error between the predicted value and the true value, and its value range is [0, +∞]; R2 represents the determination coefficient, and its value range is [0, 1].

8. A power user load forecasting system integrating RPA-AI technology using the method according to any one of claims 1 to 7, characterized in that: Based on the RPA digital integration module, the hierarchical structured framework of the software system is designed, and the computing, storage and network resources of the cloud platform are utilized. The two databases and one platform are combined through intelligent components to achieve efficient data processing. Power energy analysis module: Use RPA technology to automatically collect power energy data, verify, clean and classify the data; Multi-level load forecasting model construction module, which builds a multi-level power load forecasting model according to different forecasting time requirements; The platform workflow module executes workflow logic based on RPA technology, performs process modeling, process analysis, and process execution, and is responsible for realizing specific business functions and automated processes, including four major parts and fifteen sub-functions: power energy trading and data analysis, distributed photovoltaic operation and electricity fee calculation, comprehensive financial management and risk assessment, and enterprise operation management and market development.

9. A computer device comprising: Memory and processor; The memory stores a computer program, characterized in that when the processor executes the computer program, the steps of the power user load forecasting method integrating RPA-AI technology are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for predicting power user load by integrating RPA-AI technology are implemented.

Images